PixVerse

View saved version →

---
name: pixverse-ad-variants
description: "Ad variations (Ad Multiplier): make ordered independent edits of one supplied ad, including person, wardrobe, product and background changes while preserving source performance, timing, untargeted text and audio. Also adapt an ad for new products, audiences or languages."
---

# Ad Variants

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Start every variant from the same accepted source. Define what may change and what must survive before choosing an editing route.

For a similar ad with a new product, use the reference information already obtained by Codex
and continue with `../pixverse-product-video/SKILL.md`. Source-preserving edits below need
usable source footage. Codex chooses reference acquisition and understanding independently;
this plugin takes over the requested adaptation or edit.

For look changes that keep the motion, `../pixverse-video-restyle/SKILL.md` owns the route;
for versions of a project this plugin generated, `../pixverse-video-variants/SKILL.md` reuses
accepted assets. Read `./references/source-preserving.md` for the complete source-preserving variant recipe:
one 4–30s source, ordered reference mappings, full-look replacement, exact protected text,
timed operations, original-audio restoration and measured final checks. Keep its protected
properties in every prompt; a visually similar remake is not an equivalent result.

## Make A Change Sheet

Inspect the full supplied video, its duration, aspect, cuts, soundtrack, speech and visible text. For each requested variant record the source, one target change, protected elements and acceptance checks. Preserve the user's ordering and count. An open brief can start with textual options; do not generate an unrequested pack.

Distinguish local editorial changes from visual regeneration:

- Title, crop, speed, grade, mix or existing-asset arrangement: use `../pixverse-video-editing/SKILL.md`.
- Subject, wardrobe or scene replacement: evaluate current modify/reference-edit capabilities through the gateway and account-aware route.
- Motion transfer: use a route explicitly supporting motion reference; do not claim it preserves the original scene, voice or edit.

A candidate generation route is not proof of exact preservation. If a requirement exceeds current model limits, explain the affected requirement and propose a smaller edit or source-preserving composition before paid work.

For Seedance 2.5 source edits, the current route is `create reference --task-type edit`
with the original video in `--videos`, `--duration auto`, `--aspect-ratio auto` and
`--quality 1080p`. Verify the offline capability through the gateway before composing
the queue. Ordinary reference generation and `create modify` are not interchangeable
with this edit route. The CLI has no Seedance audio toggle; preserve original audio at
export. Follow the concrete prompt and finishing references linked from the source recipe.

## Prompt Example

"Use the original clip's standing presenter, actions and camera framing as the reference. Change only the wall behind her to a pale blue plaster wall under the same daylight direction. Keep her face, clothes, hand actions and product unchanged. Retain the visible sequence without adding a cut."

Keep soundtrack reuse as an explicit edit requirement outside the generation prompt; the prompt alone cannot guarantee it. Each variant binds the original source, never the preceding generated variant.

## Compare And Deliver

Compare source and result at cut boundaries, difficult motions and text moments. Check full duration, frame composition, face/product fidelity and temporal behavior. Where authorized and synchronization still matches, reuse the original audio and exact overlays deterministically. Changed lip motion can invalidate reuse of the original speech.

Deliver the requested variants with a compact change sheet. Report any preservation failure instead of silently replacing “same motion/timing” with “similar style.” Repair only the rejected version.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 4

View saved version →

---
name: pixverse-audio
description: "Music and sound: create music or sound assets, or route audio work to narration and mixing. Use voiceover for a spoken delivery and captions for subtitles."
---

# Voice and Music

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Read `../../skills-shared/audio-craft.md`.

For standalone SFX/reference-audio capabilities, read
`../../skills-shared/extended-capabilities.md`. Native UGC voices follow their
specialized workflow; do not simplify its audio map into a sparse generic prompt.

## Voice

Use standalone voice when narration needs timing, reuse, subtitles, or download:

```bash
pixverse voice presets --model speech-2.8-hd --language zh --json
pixverse create voice --model speech-2.8-hd --voice-id <preset_voice_id> --text "..." --json
```

Use `pixverse voice models` and `pixverse voice presets --model <id>` when the user wants voice choice.

For natural or timed narration, use `../pixverse-voiceover/SKILL.md`. Generate natural speech sections, measure their audio and align captions afterward through `../pixverse-captions/SKILL.md`. Keep the final SRT under `projects/<slug>/prompts/` or `projects/<slug>/assets/subtitles/`. The following segmented queue is only for an explicit fixed-window take list:

```bash
"${PVX}" subtitles clean <srt> <clean.srt>
"${PVX}" subtitles inspect <clean.srt>
"${PVX}" subtitles style --format force-style
"${PVX}" subtitles voice-queue <clean.srt> projects/<slug>/voice-queue.json --project <slug> --segments-dir projects/<slug>/prompts/tts-segments --voice-id <preset_voice_id> --language zh
```

Existing continuous audio can receive accurately aligned captions without another generation. Preserve accepted spoken words; clean display punctuation only when that style is wanted. Start from the plugin subtitle style and verify readability in the actual frame.

## Music

Use standalone music when BGM/soundtrack is requested:

```bash
pixverse music models --json
pixverse create music --model music-2.6 --prompt "..." --instrumental --json
```

Use lyrics only when vocals are part of the concept. `music-3.0` (MiniMax Music 3.0)
and `music-v2` (ElevenLabs Music V2) support lyrics, auto lyrics and instrumental
generation; the default remains `music-2.6`. Read
`../../skills-shared/pixverse-cli-1.4.4.md` for the reviewed alternatives.

Music is generated at auto-duration and edited to the picture; `--duration-seconds` and
`--no-duration-auto` are refused by the queue because the service rejects fixed targets.
Treat a requested length as an edit requirement: measure the returned audio and trim/fade
locally. See the shared audio craft reference.

## Video Audio

Use in-video generated sound when the brief wants the model to synchronize ambience and sound effects to the generated motion. Do not describe this as "music none" shorthand; write the no-music requirement in plain natural language.

Before using `--audio`, decide and record one of these modes:

These switches apply only when the selected model exposes them. Seedance 2.5 omits both
`--audio` and `--no-audio`; keep sound intent in the prompt, inspect returned sound, and mute
locally for a silent deliverable. Missing toggle support does not establish missing audio.

MiniMax H3/H3 Max is a different contract: generated audio is unsupported. Its
`--audios` accepts reference inputs; it does not enable output sound. Plan separate
voice/music or local sound assets when sound is required, preserving paid approval.

| Mode | Use when | Command stance |
|---|---|---|
| synced SFX/ambience | model should create sound tied to visible motion, with no music | `--audio`; use natural-language no-music instruction |
| clean picture | sound will be rebuilt from separate stems or user asked for silent picture | `--no-audio` |
| fused native audio | casual social/UGC clip where baked audio including music may be acceptable | `--audio` where supported; preserve the workflow’s complete sound/voice plan |

For in-video generated sound, use a plain-language audio paragraph and avoid musical cue words:

```text
This video should not generate any music, background music, score, melody, rhythmic bed, or trailer-style musical hit.
Only generate synchronized sound effects and environmental ambience for what is visible in the shot.
The sound should be natural and quiet: physical Foley from the characters and objects, room tone, air, footsteps, fabric, breath, water, and other diegetic sounds.
Do not add orchestral emotion, musical chimes, melodic pads, percussion, or any soundtrack-like layer.
```

If the result contains unwanted music or unpleasant SFX/ambience, mark that audio rejected and strip/replace it. Do not mix BGM over a bad generated track.

If `quote queue` adds an audio note, treat it as an advisory reminder. If the prompt already says "no music" or the user's intent is clear, proceed to confirmation without changing the route. Only rewrite the sound paragraph when it is quick and obviously improves the prompt; do not switch to `--no-audio` or delay the run just to silence the note.

For post overlay, use `../pixverse-video-editing/SKILL.md` and keep audio assets separate.

Referenced files: 1

View saved version →

---
name: pixverse-brand-system
description: "Brand visuals: generate coherent logo concepts, palette boards, brand guide images and branded mockups using the user's accepted visual references."
---

# Brand Visuals

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Build a small set of decisions that can be reused across real deliverables. Each application should derive from the same accepted identity.

## Required Recipes

Read `./references/foundation.md` for every brand task: source assets, three distinct
logo mechanisms, palette/type direction, durable decisions and selective revisions.
Read `./references/applications.md` for mockups, packaging, merchandise, signage, social
graphics, posters and brand guide images. Preserve their concrete stage
and output contracts. Existing assets and explicit delegated choices take precedence.

## Define The Direction

Establish the name, audience, offering, personality and intended uses from the brief. Separate fixed brand facts from exploratory choices. Start with concise distinct directions only when the user needs concept selection; do not generate a large pack before the identity is decided.

Generate distinct mark and wordmark concepts as images. Keep silhouette, letterforms, spacing and small-size legibility deliberate. Supplied official logos remain authoritative references. Record the selected concept and palette for later generated applications.

## Make The System

Record palette values, readable text/background combinations, typography direction, spacing and image treatment. Use a supplied font specimen when precise letterforms matter, and distinguish the intended font from visually generated lettering.

Use a simple dependency chain: accepted name/mark, core tokens, layout templates, applications. A mark change updates the dependent guide and examples; it should not leave old logos scattered through the output.

Provide relevant image outputs: profile image, social card, packaging mockup, cover or hero visual. Use `../pixverse-product-stills/SKILL.md` for photo applications and `../pixverse-listing-images/SKILL.md` for information cards. Keep original marks as explicit edit references in mockups.

## Example Direction

For a fictional neighborhood repair studio: a compact joint motif, warm off-white field, deep blue text, restrained orange accents, sturdy humanist lettering and close photographs of tools. The motif should remain recognizable at small size.

## Deliver And Revise

Deliver the agreed PNG/JPG image set with palette values, accepted copy and source references. Check small-size legibility, contrast and consistency. Refine images through Sunburst using the accepted result; regenerate only affected applications.

Keep the delivered image count and contents aligned with the requested brand visuals.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 3

View saved version →

---
name: pixverse-canvas
description: "PixVerse Canvas: create, open, continue or edit a visual project with connected image/video nodes, references and native composition. Use for an explicitly selected Canvas workflow."
---

# PixVerse Canvas

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Use this public entry to go directly into the Canvas workflow. Studio is not a prerequisite.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two
directories above this skill directory and contains `.codex-plugin/plugin.json`. In each new shell
session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helpers through `"${PVX}"`, preserving the user's working directory. Do not assume the current
directory is the plugin checkout or bypass the installed wrapper with a module invocation.

## Direct Canvas Route

1. Confirm that Canvas is the intended work target from the user's request or selected-project
   continuation. An ambient browser tab or an old `.canvas-project.json` alone does not select this
   route. An explicit non-Canvas request keeps its ordinary workflow.
2. Read `../../skills-internal/pixverse-agent-canvas/SKILL.md` immediately. It is the single canonical
   operating contract; follow it for binding, early browser handoff, graph operations, paid-work
   authorization, cloud preview, recovery, and delivery. This entry does not define a second workflow.
3. Before the first browser action, read `../../skills-shared/web-handoff.md`. Use the host's current
   top-level Computer Use/CUA entrypoint when exposed, follow the documentation returned by its first
   call, and apply the Codex IAB-only handoff and unavailable-browser fallback. With a resolved project,
   make the early browser handoff the first visible milestone, not a final delivery step.

Do not load Studio, Production, Delivery, or the full internal root merely to reach Canvas. Load other
playbooks only for a concrete need identified by the canonical Canvas contract. If that contract is
already loaded in this turn, continue from it without repeating initialization or browser handoff.

The canonical **Cloud Preview And On-Demand Delivery** rules override ordinary queue/local defaults
only for this selected Canvas target. Opening or inspecting a project is not permission to mutate or
generate; creating media still follows the canonical preflight and effective confirmation policy.

Referenced files: 1

View saved version →

---
name: pixverse-captions
description: "Video captions: create or correct timed subtitles from actual speech and render readable captions into supplied video, including translated subtitles when requested."
---

# Captions

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Make captions follow what the audience hears. Work from the actual recording, even when a draft script exists.

For a burned-in deliverable, read `./references/burn-in.md` before transcription or layout.
It defines the four ordered gates, verified-word alignment, bold/paper/clean/UGC looks,
safe geometry, glyph coverage and voice-tail checks. Sidecar and translation requests
remain separate; neither requires rendering new video unless requested.

## Establish The Words And Clock

Inspect the video duration, audio tracks, frame size, orientation and requested output. Reuse a supplied accurate timed transcript. Otherwise transcribe or align with a tool actually available in this host, then check names, numbers, language changes and uncertain words against the audio. If alignment is unavailable, offer a clearly marked timing draft or manually time a short clip; do not present guessed word timestamps as measured.

A supplied script is a candidate transcript until compared with the recording. Translation may change displayed wording, but retain the original audio time windows and meaning. Keep original and translated tracks separate when both are requested.

## Compose Readable Captions

- Split by meaning and breathing, not by fixed character counts alone. Avoid flashing one-word fragments unless requested.
- Measure text with the selected font. One or two lines may be appropriate; leave faces, product actions and platform controls visible.
- Use the plugin subtitle style as a starting point, then check the actual portrait or landscape frame. Preserve user typography and punctuation preferences.
- Export a timed SRT or ASS. Exact text is a render layer; never ask the video model to draw the captions.
- For karaoke or word emphasis, use measured word timing. Sentence timing cannot justify precise word highlights.
- For word-highlighted or word-by-word captions, write the lines as a script (`../../skills-shared/semantic-script.md`),
  align them to the recording (`../../skills-shared/word-timing.md`) and generate the ASS with
  `"${PVX}" timeline captions <timeline.json> --to captions.ass --style karaoke|bold|clean|ugc|pop`;
  role colours come from the script's speaker labels. Burn with the FFmpeg `ass` filter as usual.

## Render And Check

For ordinary local delivery, use FFmpeg and the existing subtitle helpers in `../pixverse-video-editing/SKILL.md`. Test font availability and renderer support before a full export. Inspect representative frames with short, long and mixed-language captions; watch the start and end plus several sentence transitions.

Preserve the selected soundtrack and requested video properties. Verify duration, readable placement, encoding and caption sync after rendering. A subtitle file beside an uncaptioned video does not satisfy burned-in delivery.

Changing color, font or line breaks only rerenders captions. Do not regenerate the video or TTS. Deliver the captioned file and the editable timing file; retain the source video.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-character-sheet
description: "Character design: create reusable character sheets, expressions, turnarounds and identity references for consistent scenes, animation or episodic video."
---

# Character Sheets

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Use this when a person, creature, mascot, prop, outfit, or vehicle must survive across multiple images/videos.

Read:

- `../../skills-shared/reference-analysis.md`
- `../../skills-shared/prompt-craft.md`

Read `./references/identity.md` for accepted reference images, supported view evidence
and consistent identity, wardrobe and scene roles across generated shots.

## Fast Path

1. Collect existing references or write identity anchors.
2. Generate character/item sheet before narrative clips.
3. Record sheet task id and URL in project memory.
4. Use reference mode or image-to-video for future clips.

## Character Sheet Prompt

Include:

- front, side, back, head detail
- outfit materials and colors
- silhouette
- age/body/posture
- face/hair/accessories
- expression range
- neutral clean background

## Prop Sheet Prompt

Include:

- front/side/top views
- material
- scale relationship
- logo/mark placement
- wear/detail
- what must never change

## CLI Shape

```bash
pixverse create image --model gpt-image-2.5-sunburst --quality 1440p --detail-level high --aspect-ratio 16:9 --images <refs...> --prompt "..."
```

Then:

```bash
pixverse create reference --images <character-sheet> <prop-sheet> --prompt "..."
```

## Quality Checks

- all views describe the same character/item
- no extra unwanted characters
- face and outfit are stable enough
- sheet is useful as future reference
- memory maps name -> task id, absolute local path, provider URL, and role

Referenced files: 2

View saved version →

---
name: pixverse-cinematic-story
description: "Story clip: turn a compact narrative into a few connected shots with a clear visual beat, controlled pacing and coherent sound."
---

# Cinematic Story Clips

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Tell one beat well. A 15-second clip cannot carry a feature plot.

Read:

- `../../skills-shared/cinematography.md`
- `../../skills-shared/prompt-craft.md`
- `../../skills-shared/model-routing.md`

## Fast Path

1. Compress the idea into one emotional turn.
2. Define character, want, obstacle, image hook, ending image.
3. For a simple one-shot story, generate video directly. Generate a board only when character/world/composition continuity makes it useful and the user approves that paid stage.
4. Render the clip.
5. QA for continuity and emotional readability.

For a multi-shot story, establish character and scene/look assets, write the detailed storyboard, and create one approved control frame per shot. Once those are approved, batch the motion stage:

```bash
"${PVX}" story queue projects/<slug>/story.json --project <slug> --shot <shot-1> --shot <shot-2> --shot <shot-3> --shot-reference <shot-1-board> --shot-reference <shot-2-board> --shot-reference <shot-3-board> --reference <character-or-world-lock> --target-duration 30 --music-prompt <instrumental-score-prompt-or-file> --audio --preflight --format markdown
```

When the story needs a reusable score, `--music-prompt` adds one independent Music 2.6 instrumental task to the same queue. It starts alongside the board and removes the extra `queue append` preparation step; trim and mix it locally after `story assemble`.

Never use one contact sheet, storyboard grid, or multi-panel sheet as the direct reference for every Seedance shot. Each shot-specific reference must be a clean cinematic frame for that shot. The compact shared `--board-prompt` route is opt-in only for low-risk stories with no visual-selection checkpoint.

After generation, use `"${PVX}" story assemble <slug> --output projects/<slug>/deliverables/final-story.mp4 --sample-frames`. This downloads missing shots, stitches, checks duration, samples across the full timeline, and writes project memory in one local call.

## Story Shapes

- **Reveal:** ordinary image becomes strange.
- **Choice:** character hesitates, acts, consequence lands.
- **Chase:** motion pressure, one clean geography.
- **Gag:** setup, escalation, reversal.
- **Trailer:** three iconic fragments, final title/feeling.
- **Fable:** symbolic character, clear moral image.

## Prompt Contract

For video, include:

- protagonist identity
- setting
- action timeline
- camera language
- emotional turn
- ending frame
- audio/dialogue only if necessary

Avoid cramming 6 scenes into one prompt. If the idea needs multiple clips, create multiple tasks and later stitch.

## Quality Checks

- can the viewer infer what changed?
- does the protagonist stay recognizable?
- does camera motion help, not confuse?
- does the ending frame feel intentional?
- is the clip too generic for the user's taste?

Referenced files: 1

View saved version →

---
name: pixverse-cover-art
description: "Video thumbnail and YouTube cover: create a clear, compelling cover with controlled subject identity, mobile-readable composition and editable exact text."
---

# Video Covers

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


A cover makes one understandable promise about the video. Find that promise before decorating the frame.

Read `./references/thumbnail-recipes.md` before concept selection or prompts. It supplies
16 frameworks, at least five internal concept candidates, exact identity/reference roles,
the eleven-block prompt, text/split gates, emotion/take variation and surgical edits.
Default to a clean image; a headline or people-led concept needs the corresponding resolved
input. Do not pass a style-only thumbnail as an identity reference.

## Choose The Idea

Read the title or content summary and inspect supplied faces/products. Identify the evidence object or visible contrast that explains why someone would watch. Choose one central relationship: person and result, small and large, before and after, question and evidence, or a revealing detail.

Use a small number of concept descriptions when direction is open. Generate only the number of candidates requested or included in the approved plan. Do not invent sensational events or results absent from the content.

## Build The Image

1. Place the face or key object large enough to read at thumbnail size; preserve reference identity.
2. Limit competing elements. Use lighting, crop and contrast to direct attention.
3. Generate the required image plate. Keep title words and exact brand marks in a controlled text/vector layer when precision matters.
4. Export at the requested platform ratio. Review both full resolution and a roughly 120–160 pixel-wide preview.

## Prompt Example

“Preserve the supplied presenter’s face and glasses. She holds the real pocket projector at chest height, looking toward its small sharp projection on the wall. Waist-up framing on the right, a dark uncluttered wall on the left reserved for the exact title. Plausible warm room lighting, curious expression rather than exaggerated shock. Keep projector lens and buttons unchanged.”

## Check The Promise

Can a small image communicate the main relationship? Is the title readable and truthful? Are the face, product and essential detail intact? Deliver the finished image; a blank text plate alone is incomplete. Changing the title should usually only require layout and export. Do not claim improved click-through without an actual experiment.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-create-image
description: "Image generation: create or edit a still image with PixVerse, preserving requested references, model and format. Use for a direct image deliverable."
---

# Create Images

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Use image generation for standalone stills and as a control layer before video.

For a simple standalone image, use the shared contract's account-aware `route queue`
fast path directly. Read `../../skills-shared/prompt-craft.md` for complex composition
or prompt repair; read `../../skills-shared/model-routing.md` for a custom model,
unfamiliar parameters or route comparison. The supported default needs neither lookup.
Read `../../skills-shared/capability-priority.md` when the image is part of a video pipeline.

For reference roles, image edits and reusable product/brand inputs, read
`../../skills-shared/extended-capabilities.md` before composing the request.

## Common Outputs

- visual territory boards
- product hero stills
- storyboard sheets
- character sheets
- item/prop references
- locked logo/title/final-frame plates
- posters, KVs, book covers, album covers
- game/UI mock screens

## Default Command

Use the same high-quality default for final stills, previews and intermediate control art:

```bash
pixverse create image --model gpt-image-2.5-sunburst --quality 1440p --detail-level high --aspect-ratio <ratio> --prompt "..."
```

Here `1440p` is the CLI's 2K tier, while `high` controls rendering detail. Do not reduce
control boards to 1080p/medium. Respect an explicit user override.

Free/Basic or model-entitlement failure pauses for the subscription/fallback choice in
`../../skills-shared/quality-policy.md`. Only after the user accepts fallback:

```bash
pixverse create image --model gemini-3.1-flash-lite --quality 1080p --aspect-ratio <ratio> --prompt "..."
```

Nano Banana 2 Lite does not accept GPT's detail-level flag. Other image families remain
available when explicitly selected; verify their own supported parameters. Read
`../../skills-shared/pixverse-cli-1.4.4.md` for Flare/Sunburst's expanded detail/ratio
choices and Seedream's model-specific quality and reference limits. Use a manual
queue for explicit models outside the automatic route helper's image families.

Use uploaded images with `--image` or `--images` when they define the subject, product, character, logo, or layout.

## Image Before Video

Prefer image-first when:

- subject identity matters
- product packaging matters
- text/logo must land exactly enough to inspect
- the user wants multiple concepts
- video would be expensive to explore blindly
- a storyboard frame, title card, final packshot, UI plate, or character sheet will control later video

## Video Control Roles

Label stills by role before using them downstream:

- identity/product anchor
- storyboard opening frame
- style reference
- exact text/logo/UI plate
- prop/wardrobe/vehicle sheet
- negative example or rejected route

Do not send an asset sheet as an I2V opening frame when a proper storyboard frame is needed; generate or choose the shot frame first.

## QA When Requested

Standalone images deliver directly after generation and download, with no automatic visual inspection or QA command. If the user requests review or a specialized workflow explicitly requires an anchor check, assess only the requested fit, crop, subject, text or logo constraints.

Referenced files: 1

View saved version →

---
name: pixverse-create-video
description: "Video generation: create a clip from a prompt or reference through PixVerse. Handle requested duration, motion and sound; simple clips start directly with video."
---

# Create Videos

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Read this when the user wants a moving-image result. Ordinary single clips deliver directly after generation and download, without automatic playback inspection, frame sampling, audio checks or QA commands. Checks below apply only when a selected specialized production workflow explicitly needs them or the user requests review.

For a simple prompt-only clip, use the shared contract's direct `route queue` fast path.
Load these references only when the brief needs them:

- `../../skills-shared/capability-priority.md`: exact control assets, motion/reference modes
- `../../skills-shared/model-routing.md`: custom models, unfamiliar parameters or comparison
- `../../skills-shared/prompt-craft.md`: complex action, composition or prompt repair

For sketch-at-timestamp edits, generative reframing, reference-driven ad inputs or
other non-model workflows, read `../../skills-shared/extended-capabilities.md`.
Preserve a selected specialist workflow’s shot/board/voice recipe over generic
production-depth advice below.

## Choose Route

For an explicitly requested MiniMax H3 Max or other alternate model, read
`../../skills-shared/pixverse-cli-1.4.4.md`, then author a capability-checked manual
queue. H3 Max supports video/reference/two-frame transition, with distinct framing
rules and at most 12 mixed references; it has no generated-audio control.

Use Seedance 2.5 at 1080p for all supported video modes. Apply `../../skills-internal/pixverse-seedance-prompt-enhance/SKILL.md` before preflight. Free/Basic must stop for the subscription/fallback choice; only an accepted fallback uses v6 540p. Preserve the requested mode and references, and explain any capability conflict.

| User intent | Route |
|---|---|
| single simple lightweight prompt-only clip | direct video; Seedance 2.5 1080p; v6 540p only after fallback consent; no hidden image task |
| serious/high-fidelity single clip without identity lock | Seedance 2.5 1080p; v6 540p only after fallback consent |
| short clip with exact character/product/opening-frame lock | explicitly approved account-compatible image/plate, then v6 or Seedance 2.5 video/reference |
| style/identity uncertain | image board first, then video |
| user uploaded subject/product/character | image-to-video or Seedance reference; do not drop the asset |
| multiple subjects need consistency | `pixverse create reference` |
| first/last frame transformation | `pixverse create transition` |
| motion follows a reference video | Seedance `create reference --task-type reference`; assign the video's motion role |
| continue an existing clip | Seedance reference `--task-type extend` |
| change an existing clip | Seedance reference `--task-type edit` |
| improve resolution | `pixverse create upscale` |
| stitch clips/subtitle/BGM | `../pixverse-video-editing/SKILL.md` |

## Pick Production Skill

- product/brand/ecommerce: `../pixverse-product-video/SKILL.md`
- creator/social ad: `../pixverse-ugc-video/SKILL.md`
- music/dance/lyrics/performance: `../pixverse-music-video/SKILL.md`
- narrative/trailer/story scene: `../pixverse-cinematic-story/SKILL.md`
- vertical short drama: `../pixverse-short-drama/SKILL.md`
- game teaser: `../pixverse-game-trailer/SKILL.md`
- recurring cast/props: `../pixverse-character-sheet/SKILL.md`

## Default Work Shape

Classify dynamically before routing:

- Quick task: usually one shot/beat, no recurring identity, no exact plate, low revision stakes. Generate the requested medium directly.
- Controlled shot: short can still be serious when product, character, exact first frame, UI/text, or brand fidelity matters. Add only the necessary explicit control asset.
- Short-film project: multiple scenes/beats, recurring cast, animation, dialogue, trailer/advertising delivery, or revision workflow. Lock character, scene/look, and individual shot boards before motion.
- Long-form project: episodic or feature-length. Lock script and bibles, then divide into acts/reels/scenes/shots and independently approved scene batches.

For serious work:

1. clarify only if a missing choice changes cost or format
2. classify user assets as identity, product, storyboard, style, motion, audio, or raw material
3. plan 2-4 creative routes when aesthetics are open
4. generate still control assets only after their own quote and approval
5. show and approve character/look assets before per-shot boards; show boards before video
6. render video from the approved per-shot route
7. surface each successful preview immediately; perform QA only when this specialized workflow names required checks, then record memory and hand off project files

## Model Bias

- Use `seedance-2.5` at 1080p for drafts and final video, with the prompt-enhancement skill.
- Use its `video`, `reference` or two-frame `transition` contract as appropriate. Edits and
  extensions use `create reference --task-type edit|extend`; motion references also use reference.
- Preserve the reference chain on creative failure. Account failures stop for user choice.
- After explicit fallback consent, use `v6` at 540p only within its supported modes/limits.
- Seedance 2.5 has no CLI audio, multi-shot or off-peak flags. Put sound/cuts in the prompt.
  A guaranteed silent export requires removing its audio track through the editing workflow.
- Ordinary generation still delivers directly; enhancement does not add automatic media QA.

Read `../../skills-shared/exploration-patterns.md` when the user wants "cool", "surprising", "not generic", or "like this reference".

Referenced files: 1

View saved version →

---
name: pixverse-delivery
description: "Video delivery: recover PixVerse results, download existing assets, continue a project or package final files. Run media QA when explicitly requested."
---

# PixVerse Delivery

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Use this to recover, inspect or export an existing project. Ordinary image/video generation completes directly in its creation workflow; it does not need this skill or QA. In this skill, load quality playbooks and perform inspection only when the user requests QA or a specialized workflow explicitly requires named checks. Editing requests enter `../pixverse-video-editing/SKILL.md` directly; production workflows own their final handoff.

Before any new paid generation or retry, follow `../../skills-shared/cli-workflow.md`.

For an explicitly selected Canvas target, use `../pixverse-canvas/SKILL.md` as the direct entry. If its
internal Canvas playbook is already loaded, continue under it without restarting the handoff or
loading ordinary local-delivery playbooks merely because this entry was also selected.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Internal Playbooks

- project memory, preferences, useful decisions, asset mapping, rejections: `../../skills-internal/pixverse-agent-project-memory/SKILL.md`
- QA, inspection, retry logic, final delivery checks: `../../skills-internal/pixverse-agent-quality/SKILL.md`
- stitching, subtitles, BGM, narration overlay, trimming, packaging: `../pixverse-video-editing/SKILL.md`
- asset info, task ids, downloads, CLI syntax: `../../skills-internal/pixverse-agent-gateway/SKILL.md`
- Canvas-native editing, graph reuse, composition, compression, and guarded mutation: `../../skills-internal/pixverse-agent-canvas/SKILL.md`

## Canvas Delivery Routing

Apply this routing only when Canvas has been selected as the current target under the Canvas
playbook's activation boundary. Then prefer current executable capability-declared Canvas nodes such as
`video_compose`, reuse the existing graph assets, and avoid
download/edit/re-upload loops. Local finishing remains available for requirements the Canvas schema
cannot express, an explicitly requested local master, or final delivery-only packaging. If a local
fallback must return to Canvas, upload only its final result once.
`video_compress` is currently a reserved, non-executable, non-authorable compatibility type and must
not be constructed or dispatched.

For every non-Canvas source project or ordinary queue delivery, skip Canvas capability discovery and
retain the existing local preview, download, FFmpeg, subtitle, QA, assembly, and packaging workflow.
An unrelated or historical Canvas binding does not change that route.

For the explicit Canvas target, follow the Canvas skill's **Cloud Preview And On-Demand Delivery**
contract throughout delivery and recovery. Preview in Canvas; do not automatically download even
intermediate assets, run local QA, refresh the balance, or assemble an invoice. Follow cloud status
and reuse node/asset references. Use `project resume [hint] --surface canvas --format markdown` and
`project handoff <slug> --surface canvas --format markdown`, not queue-only `qa project` or ledger
recovery. Report stage, cloud nodes, remaining work, and unverified technical properties. If IAB is
unavailable, return the Canvas link without silently downloading or claiming visual QA passed.

Automatic Canvas review also follows **Preserve The User's Viewing State** in that playbook: no
enlarged dialogs, zoom, fullscreen, or maximization without the user's explicit enlargement request.
Preserve an already enlarged user view; failed controls leave checks unperformed, not permission to
change the viewing mode.

An explicit request for local files enables only the required export/download and targeted file QA;
credit reporting remains a separate opt-in. Record a completed file handoff with
`--delivery-mode local --deliverable-path <actual-file>`; cloud completion does not need local media.

## Ordinary Queue And Local Production Delivery Contract

The remaining queue/ledger workflow is for non-Canvas production. For requested Canvas file delivery,
apply only its relevant local file checks; keep the Canvas follow/recovery contract above.

`queue run` downloads every successful generated asset under `projects/<slug>/assets/` and returns an absolute `local_path`. As soon as that path exists, show the primary image/video from the local file using the host's local-media syntax. Do not use the provider URL as the Codex inline preview source; it remains useful only for audit or external sharing. Do not hold the local preview behind deep QA, billing reconciliation, ledger recovery, packaging, or analysis. Use the returned result for ordinary delivery; only perform QA afterward when requested or explicitly required by a specialized workflow.

If `local_preview_status` is not `ready`, say that generation succeeded but local delivery failed. Retry the free `pixverse asset download` path by task id; do not regenerate and do not fall back to a remote URL for Codex display.

When continuing an existing project, read memory before acting:

```bash
"${PVX}" project resume [title-or-slug-hint] --format markdown
"${PVX}" project portfolio --format markdown
"${PVX}" project search <slug> "<asset, task id, style, rejection, or cue>"
"${PVX}" project ledger <slug> --format markdown
```

Use `project resume` first for “continue”, “the earlier one”, or a fuzzy title. It returns the likely project, stage, compact memory, latest assets, unresolved work, and one recommended next action without dumping provider payloads. Use `project portfolio` when multiple projects may match; reserve `project summary` for deep raw trace inspection.

When QA is requested or a specialized workflow explicitly requires inspection, verify the relevant items:

- promised generated assets exist as local files; a provider URL alone is not a completed Codex delivery
- media type, aspect ratio, duration, and audio match the brief
- product, character, UI, or text constraints survived well enough
- the result obeys the user's intended aesthetic, including deliberate roughness, historical texture, or imperfect capture
- no required reference, product plate, storyboard frame, or control mode was silently dropped
- project memory records useful task ids, absolute local preview paths, provider URLs, and creative decisions
- `project ledger <slug> --format markdown` or `queue run` output gives the final invoice-style asset ledger
- retry recommendations are cost-aware and explicit

For local inspection:

```bash
"${PVX}" qa project <slug>
"${PVX}" qa inspect <file-or-url> --project <slug>
"${PVX}" qa inspect <video> --project <slug> --expect-duration <seconds> --expect-aspect-ratio <ratio>
```

When project QA is requested, prefer `qa project <slug>` for the latest completed run so technical checks, the generated-asset ledger, and the invoice are recovered together instead of one asset per host-tool call.
When whole-project QA is requested after several generation batches, use `qa project <slug> --all-runs` to inspect every unique historical asset in one call; idempotent reruns are deduplicated while each asset keeps the expectations from its own quote.

For a non-Canvas story queue made of `shot-01`, `shot-02`, and later shots, close the local post-production loop in one call:

```bash
"${PVX}" story assemble <slug> --output projects/<slug>/deliverables/final-story.mp4 --sample-frames
```

It reads ordered story assets from the latest ledger, reuses local downloads or downloads missing clips by task id, stitches them through FFmpeg, checks the combined duration, runs technical QA, samples the full timeline when requested, and records the deliverable in project memory. Pass ordered `--asset-id` values for an older run or explicit ordered `--clip` paths for a custom cut.

For subtitles and narration, use actual audio as the clock and an aligned SRT/ASS for rendering. Follow `../pixverse-captions/SKILL.md`; a caption-only edit does not regenerate voice. Use `../pixverse-voiceover/SKILL.md` and the segmented voice-queue helper below only for an explicit fixed-window take plan. Preserve spoken text and honor the requested display typography:

```bash
"${PVX}" subtitles clean projects/<slug>/prompts/narration.srt projects/<slug>/prompts/narration.clean.srt
"${PVX}" subtitles inspect projects/<slug>/prompts/narration.clean.srt
"${PVX}" subtitles style --format force-style
"${PVX}" subtitles voice-queue projects/<slug>/prompts/narration.clean.srt projects/<slug>/voice-queue.json --project <slug> --segments-dir projects/<slug>/prompts/tts-segments --voice-id <voice_id> --language zh --speed 0.9
```

Do not end displayed subtitle lines with sentence punctuation unless the user explicitly asks for that typography.

Keep generated media, manifests, quality reports, and notes under `projects/<slug>/`.

At the end of every meaningful stage—not only final delivery—show the project handoff:

```bash
"${PVX}" project handoff <slug> --stage <planning|character-lock|storyboard|control-assets|preview|render|final> --format markdown
```

This makes editable plans, prompts, queues, project memory, QA/audit files, generated media, and deliverables visible instead of hiding project management inside the plugin.

Referenced files: 1

View saved version →

---
name: pixverse-explainer
description: "Explainer video: turn a topic, article or script into a complete narrated film, including editorial documentary motion graphics, animated explanations, history, kids and picture stories, in any consistent non-photoreal look."
---

# Explainer Videos

A faceless narrated film is a fast, concrete story told twice at once: the narrator
says it and the picture acts it out, phrase by phrase, in one locked look. Run the
phases below in order. Each phase names the one file that holds its templates; open it
when the phase says so and fill the templates in rather than inventing new structure.

## Golden Rules

1. **Story, not essay.** People, objects, dates and numbers doing things. "A silk
   merchant ordered a portrait of his wife. Leonardo never delivered it." beats
   "Fame can grow from an absence." Abstract nouns belong in neither voice nor picture.
2. **Every spoken phrase gets a visible verb.** While the narrator says it, something
   on screen does it: a figure walks off with the panel, glaze sheets stack, a hand
   lifts the painting out of its frame, a crowd grows row by row.
3. **A shot is a process, never a pose.** Motion starts on frame one and the picture
   reaches a new state about every second. Elements enter staggered, land with weight,
   and the next thing is already arriving. Nothing sits still for a full second.
4. **Fill the frame.** The hero subject occupies a third to two thirds of frame
   height. Small props lost in empty paper read as a slide deck.
5. **Use actors.** Figures in the film's look perform the story with their bodies,
   mouths closed; the narrator does the talking. A giant hand, a stamp, a crowd are
   actors too. Diagrams appear when the story reaches a quantity or comparison.
6. **One look, pasted not paraphrased.** Take the selected style's FORMULA from its
   recipe file and paste it byte-identical into every image and video prompt. Style
   words live only there; shot lines are pure choreography.
7. **Compose from the roster.** Every block references its stage, cast and props
   images. Never animate from the style key alone.
8. **Prompts say what happens.** Write positive, visible instructions. Keep exclusions
   to the single NEGATIVE line of the block template. No commentary, no reasons, no
   "this shows that…" inside a prompt.
9. **Narration is brisk and fills the film.** Natural TTS speed (0.95–1.05, never a
   slowed read), each block's line filling 8.2–9.6s of its 10s window.
10. **Words on screen come from the edit.** Generated video carries no letters or
    digits; captions and exact labels are burned in afterwards.

## Pipeline

**Phase 0 — Resolve the brief.** Topic, language, duration, aspect (16:9 default),
channel, look, captions, music, cover. Use what the user gave; pick sensible defaults
for the rest, state them in one line and continue. `./references/channels-and-formats.md`
maps channels; Kids, talking cast, song, stills and 10-minute history load
`./references/branches.md`. Model route: `./references/model-route.md`.

**Phase 1 — Script.** Open `./references/script.md`. Research, pick the through-line,
write the block lines, run its rewrite pass. Gate: every line passes the phrase→verb
test and the word budget.

**Phase 2 — Look.** `./references/style-grammar.md` indexes the recipes; open the one
selected recipe and copy its FORMULA, MOTION token and PALETTE LOCK, locking the accent
colour. Generate one style key. Gate: the key shows the look on a real subject.

**Phase 3 — Roster.** Follow the roster section of `./references/production.md`:
cast, stage plates, props, each generated with the style key attached and the FORMULA
pasted. Gate: every noun the script puts on screen has an image.

**Phase 4 — Narration.** Follow `./references/production.md`: generate one take per
block with the locked voice, measure it, and read off the phrase clock. Gate: every
take fills its window.

**Phase 5 — Block prompts and generation.** Open `./references/block-prompt.md`. Fill
its template once per 10s block with shot boundaries from the phrase clock, run its
checklist, submit all blocks in one queue.

**Phase 6 — Assemble.** Follow `./references/production.md`: check the blocks for the
named defects, assemble, mix, burn captions. Deliver one film plus caption/timing files
and sources.

Prompt-only requests deliver the script with its phrase→verb tables, roster prompts,
block prompts on an even shot grid and estimated caption cues in one document, with no
generation and without the Execution section below. A narrow request for one still or one clip uses the
simpler direct route.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles and revisions.
For media execution follow `../../skills-shared/cli-workflow.md`; load the gateway before
manual CLI commands or queue specs. Resolve the plugin root from this skill's installed
path: set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"`. The block
template in this skill already is the finished Seedance prompt; pass it through prompt
enhancement unchanged.

Referenced files: 15

View saved version →

---
name: pixverse-fashion-video
description: "Try-on video: show clothing or wearable accessories on a consistent person, with fit, material and movement tied to supplied references."
---

# Fashion Video

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Show how a specific item sits and moves on the same wearer. For the standard creator try-on,
read `./references/production.md` before designing boards or video prompts. Its eight-shot
structure, garment states, speech split and timing are part of this workflow's result.

## Plan The Wear

Inspect the wearable reference: front/back, pattern, seams, length, closure, logo and material. Inspect the chosen wearer reference or establish one consistent adult creator. Do not infer unseen construction, body measurements or fit guarantees.

Plan 4–15 second clips with eight active beats per board: pre-wear, wearing, front pose,
hand-free texture, turn, different hand-free detail, styled pose in another room, final look.
The bag/pre-wear outfit exists only in board 1 slot 1; a hard cut performs the change, with
no dressing or bag extraction on camera. Slot 2 includes one natural twirl. Later boards
follow home tour, outdoor, home reflection and continuation roles from the reference.

Use a sequential 21:9 eight-panel board and its required realism refinement, then condition
video on the clean board, accepted creator and product. Cuts 1/2/3/5/7/8 use visible native
speech; 4/6 use the same voice over hand-free macros. Write all eight timed Cut sections
and seven hard-cut boundaries, except the reference's narrowly defined camera Pick-Up.

## Timing And Overrides

For 12 seconds the cut lengths are 1.5 / 1.75 / 1.5 / 1.25 / 1.5 / 1.25 / 1.75 / 1.5.
The reference contains the other durations, body/hand staging, loop and quality rules.
Keep a fixed camera alive with the person's movement; do not default to an uncut slow walk.
Explicit single-shot editorial, silent, location or director instructions can override this
standard UGC treatment. Preserve that decision rather than calling the adapted result the
unchanged eight-beat recipe. Do not invent a real creator's ownership or experience.

## Finish

Cut around complete actions rather than hiding every difficult moment. Add exact size/material copy as a deterministic layer only when supplied. Check face continuity, pattern, hem, fastener, body intersections and realistic hand contact. Deliver the finished video and keep accepted shots when repairing one garment defect.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-game-trailer
description: "Game trailer: turn a game brief and available assets into a paced trailer with genre cues, gameplay presentation, title beats and sound."
---

# Game Trailers

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

A game trailer sells the promise of play. It should show verbs, not just scenery.

Read:

- `../../skills-shared/cinematography.md`
- `../../skills-shared/exploration-patterns.md`
- `../../skills-shared/model-routing.md`

## Fast Path

1. Identify game genre, player fantasy, and core verbs.
2. Choose trailer mode.
3. Generate board or key art.
4. Render 8-15s teaser.
5. QA for game readability and motion energy.

Use `seedance-2.5` at 1080p for both preview and final shots. Apply `../../skills-internal/pixverse-seedance-prompt-enhance/SKILL.md` before preflight; preserve the workflow-specific creative controls.

## Trailer Modes

- **World reveal:** environment, scale, faction identity.
- **Combat showcase:** attacks, dodge, impact, enemy response.
- **Ability reveal:** one mechanic made legible.
- **Character intro:** hero silhouette, weapon, stance, power.
- **Gameplay fake:** HUD/camera grammar suggests play.

## Prompt Anchors

Use game camera language:

- third-person over-shoulder
- isometric tactical
- side-scrolling action
- first-person inspection
- boss arena
- diegetic HUD

Do not overuse film-only terms if the result should feel playable.

## Quality Checks

- does it imply actual gameplay?
- are scale and objective readable?
- is UI present only when intended?
- does motion show a verb?
- does it avoid generic fantasy splash art?

Referenced files: 1

View saved version →

---
name: pixverse-howto-video
description: "Product tutorial: show a specific product in a physically plausible sequence, with clear step labels and synchronized explanation."
---

# Product Tutorials

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Teach a real operation so a viewer can repeat it. Each action must leave the product in a state that supports the next step.

Read `./references/production.md` and `../../skills-shared/ugc-production.md` before planning
the creator tutorial. The standard recipe requires four globally numbered steps per
4-15s clip, a four-panel board, mandatory refinement, persistent step headings and native
speech. Preserve these unless the user supplies a different exact procedure or explicitly
asks for a direct single-operation demonstration.

## Establish The Procedure

Use the supplied manual, approved instructions or directly observed product behavior. List the starting state, action, expected visible result and important caution for each step. Missing mechanics require source evidence; a model's plausible animation is not evidence.

Keep the procedure short enough to see. Break a complex motion at a real state boundary. Reuse the same product, creator/hands, work surface and orientation across shots. For precision, lock each shot's control frame and name the relevant button, connector, latch or fill line.

When a control frame must be the actual first frame, explicitly select image-to-video through
the gateway. General reference conditioning alone does not promise that starting state.

## Direct Each Step

Write the physical motion in order: which hand touches which part, in which direction, and when the expected result appears. Use close framing where the user must see a detail. Avoid a face blocking the procedure.

Example: “The reference lamp stands upright with its hinge at the supplied angle. The creator's right index finger presses and releases the single circular button on the base once; the lamp illuminates after the release. The left hand stays clear. Locked close-up showing finger, button and light head together. Keep the button position and base geometry unchanged. A quiet click and normal room ambience.”

Treat that action as an example only; do not assign these mechanics to a different product.

## Explain And Deliver

Write exact `Step N — Heading` labels (or the project's explicit language) into the board,
preserve them through refinement, and retain the matching label for the entire video cut.
Use the reference's em-dash typography, consistent font/position and global numbering.
Optional hook/subtitle overlays are separate. Repair illegible labels with a controlled
compositor; never quietly remove instructional text. Match narration to visible action.

At 12s, use four 3s steps. Reserve the final 0.5-1s of the last cut for a brief CTA unless
the brief excludes it; do not add a fifth step. The production reference defines caption
geometry, transitions, mechanics, reference order and specialized checks.

Check order, contact, actual operation, state continuity and label text. If the model cannot show a necessary mechanism accurately, use verified real footage or a clear controlled diagram and describe the limitation. Deliver a finished video whose instructions remain understandable at normal speed.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-listing-images
description: "Listing graphics: create a coordinated set of product cards with accurate specifications, exact editable text and consistent product identity."
---

# Listing Images

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Turn verified product information into a readable image sequence. The image sequence should answer the buyer's questions in a useful order.

Read `./references/listing-bundles.md` for main / product-images / A+ / full-set requests.
Their default counts are 1 / 6 / 8 / 13, with exact ordered module roles; custom subsets
keep the user's list. Reuse an accepted main image. Check current official marketplace
rules when the user requests a submission-ready listing.

## Assemble The Page

- Extract item name, approved claims, specifications, variants and usage from supplied material. Missing numbers stay missing; do not invent performance, price or customer experience.
- Assign a purpose to each requested card: identification, visible feature, dimensions, comparison, operation or care. Avoid repeating the same sales sentence on every card.
- Use supplied photography or `../pixverse-product-stills/SKILL.md` for missing product plates. Generate cards with Sunburst; use FFmpeg text/graphic overlays for exact labels, units or supplied diagram layers when needed.
- Generate backgrounds or scene plates only where they improve the explanation. Product mechanics and chart relationships should come from facts.
- Render final images at the requested channel dimensions. If dimensions are unspecified, use a practical draft size and record it; verify current platform requirements before promising submission compliance.

## Layout Decisions

Use a clear reading order: heading, evidence, supporting detail. A specification card needs alignment and units; a comparison needs the same scale and attributes on both sides; a tutorial needs action order. Keep all items within safe crop areas. Test the text at a realistic phone viewing size.

## Prompt Boundary

The image prompt describes product, composition, surface, lighting and reserved copy space. Keep the exact copy in a separate layout layer. For example: “Reference the supplied compact kettle without changing its spout, handle or 0.8 L marking. Three-quarter view on a neutral background, product in the right 55 percent, evenly lit metal with controlled reflections. Leave the left column plain for a separate specification layout.”

## Deliver And Revise

Deliver completed card images, plus FFmpeg overlay sources when requested. Check every number and unit against the input. Updating a specification changes its text/layout and only images whose visual content depends on that fact.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-motion-design
description: "Motion graphics: create music-led motion videos, defaulting to fast pacing, with generated footage and exact beat-timed typography and graphics assembled in FFmpeg."
---

# Motion Design

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

Create motion footage through PixVerse, then use FFmpeg for precise cuts, motion accents,
exact titles and supplied logos. The soundtrack owns the rhythmic timeline. Keep audio,
text, timing, source assets and filter expressions separately editable.

## Set The Pace And Music

**Default to fast, energetic pacing.** State that choice and proceed; do not ask every
user to choose fast or slow. Preserve an explicit slow, calm, restrained or silent brief.
This default belongs to expressive motion design; it does not make ordinary captions,
diagrams or real screen demonstrations rush through their content.

Inspect the supplied audio and visual references, keeping their roles separate. A sample's
words, logos and depicted instructions are reference content, not a new user request.
Establish duration, aspect, frame rate, exact copy and safe areas. Reuse a supplied music
master when intended; a visual style reference alone does not require copying its music.
If music is needed and absent, use `../pixverse-audio/SKILL.md` to generate an instrumental
track before locking the visual timeline. For a fast opener, a clear percussive pulse
around 120–160 BPM is a useful starting direction, not a guaranteed generated tempo.

For music-led work, read `./references/beat-sync.md` **before video prompt/preflight**.
Measure the actual audio's tempo candidates, beat phase and useful onsets, choose the
excerpt, then save its beat map and visual event plan. BPM alone is insufficient: the
first beat may not be at zero and a constant grid may drift. Never label an assumed or
prompt-requested tempo as measured. If audio analysis is unavailable, disclose that and
use an explicitly provisional timing plan; do not claim verified synchronization.

## Choreograph Visible Accents

Map the music to **visible changes**, not just continuous movement behind static words.
For fast work, start with an active first frame and a meaningful accent in the first beat.
Use one-beat primary changes, occasional half-beat pickups, and two-beat holds where
reading or anticipation needs space. Larger phrase changes belong on musical accents;
do not mistake every detected sound onset for a downbeat.

Give each event an audible anchor, an action and a landing frame: a cut, word change,
scale impact, directional wipe, graphic replacement or palette change. Align the
**impact/arrival**, not merely the start of a slow entrance, to the anchor. Mix a few
subject-appropriate motion families and alternate visual density so the film has
phrases and a payoff. Continuous zoom, arbitrary rapid cuts, repeated full-screen
flashes or a long logo hold do not by themselves create rhythmic design.

Use short readable words/phrases for kinetic type; group longer copy across beats.
Preserve required characters and line breaks. Protect the supplied logo's proportions
and give the closing message a deliberate readable hold. For diagrams, explanation
order and accurate relationships govern the pacing. Follow an explicit slow choice by
using longer musical phrases and gentler landings while retaining sound alignment.

## Generate Ingredients, Then Set Exact Timing

Read `./references/timeline-craft.md` for FFmpeg timing and layers. Plan the smallest
useful set of source shots/layers from the event plan. A short abstract or typographic
opener need not create a paid image per beat; recurring subjects and exact visual locks
still need their appropriate control assets. Use Sunburst for needed still layers and
Seedance for motion inside a scene. Independent parts need separate assets; a flat
poster does not provide them automatically.

Describe concrete actions, shot changes and useful source handles in the enhanced
video prompt. A request to "follow 140 BPM" does not guarantee exact frame timing.
Use supported audio references only with their required visual inputs and actual
gateway limits. Finish exact cuts, typography, masks and accent timing locally against
the accepted music master. Replace unused generated audio rather than mixing it over
the master. Verify unfamiliar FFmpeg filters, fonts and encoder before rendering.

Use `../pixverse-captions/SKILL.md` for ordinary burned subtitles.

## Check The Rhythm And Deliver

This workflow requires a focused check of the rendered edit: compare selected audible
anchors with actual visual landings near the beginning, middle and end; inspect the
densest typography, first/last frames, audio joins and final encoding. Follow the
specific checks in `./references/beat-sync.md`; a populated timing file alone is not
proof of a synchronized render. Inspect a short difficult range before a complex export.

Show the first usable local preview promptly. Deliver the encoded video and retain the
music excerpt, beat map, edit plan and FFmpeg command/filter graph for revisions; provide
the edit files when requested. Distinguish measured checks from aesthetic judgment.
Fix timing, crop, color and copy locally without regenerating accepted media. New paid
work still follows the shared preflight and effective confirmation policy.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 3

View saved version →

---
name: pixverse-music-video
description: "Music video (MV): plan and assemble performance, mood or narrative shots around supplied or requested music, with deliberate rhythm and visual continuity."
---

# Music Videos

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

The music owns the edit. Visuals should have a repeatable grammar, not random pretty shots.

Read:

- `../../skills-shared/audio-craft.md`
- `../../skills-shared/cinematography.md`
- `../../skills-shared/exploration-patterns.md`
- `../../skills-shared/model-routing.md`

## Fast Path

1. Identify track mood, tempo, lyric theme, and format.
2. Choose MV mode.
3. Generate a visual board.
4. Render a 8-15s loop or teaser.
5. If music is needed, generate it as a separate asset.

## MV Modes

- **Performance:** singer/band/dancer as anchor.
- **Lyric world:** visual metaphors follow the lyrics.
- **Dance:** choreography and body motion lead.
- **Visual album:** abstract mood, texture, symbols.
- **Narrative fragment:** one emotional beat, not a full plot.
- **Loop:** seamless social visualizer.

## Prompt Anchors

Include:

- tempo feel
- color system
- performance or no-performance rule
- camera rhythm
- recurring motif
- audio relationship

## CLI Shape

Music:

```bash
pixverse create music --model music-2.6 --prompt "..." --instrumental --json
```

Default to auto-duration music. Cut, loop, fade and duck locally to the edit length.
For an explicitly requested generation target or another music model, follow the
current model contract in `../../skills-shared/audio-craft.md`; measure the returned
track before setting the picture timing.

Video:

```bash
pixverse create video --model seedance-2.5 --quality 1080p --duration 15 --aspect-ratio 9:16 --prompt "..."
```

Use `seedance-2.5` at 1080p for both preview and final shots. Apply `../../skills-internal/pixverse-seedance-prompt-enhance/SKILL.md` before preflight; preserve the workflow-specific creative controls.

## Make It Yours

Pick one visual law and hold it:

- all scenes seen through reflections
- singer never appears directly
- every cut follows a color change
- motion always moves toward camera
- lyrics become physical objects

## Quality Checks

- first frame matches track promise
- motion has rhythm
- visual motif repeats
- no random montage drift
- audio asset and video concept can actually be combined

Referenced files: 1

View saved version →

---
name: pixverse-podcast-clip
description: "Podcast clip: a two-host conversation with complementary camera views, split-screen opening, speaker-coloured captions, reactions, product handoffs and lifestyle B-roll that continues under the voice."
---

# Podcast Clip

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

Two people talk; the picture moves between their complementary views, a split-screen opening
and B-roll that covers a claim while the voice continues. Speaker identity is carried by
captions colour, seat side and voice, so the audience always knows who is talking.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Cast And Geometry

Establish the two hosts, their relationship (recommender and skeptic, expert and newcomer),
the seat sides and the shared place. The first host image defines the room and capture
language; the second is generated from it as the complementary view (opposite side, matching
distance and scale, its own background details). A supplied person is the identity reference
for their seat. Use the phone-capture kit in `../../skills-shared/prompt-kits.md`; read
`./references/conversation.md` for the view relationships and the handoff of a prop across a cut.

Give each host a short TTS voice reference (`create voice`, distinct presets) and pass it with
that host's image on every take, so the voice stays the same across the clip.

## Write The Exchange

Script with `../../skills-shared/semantic-script.md`: `HOST:` and `GUEST:` turns, cue breaks
by breath, a selection around the passage a lifestyle montage should cover (start it a beat
before the reply so the picture returns as a J-cut), a moment on the product reveal or the
punchline. Measure each segment; conversational takes of 5–8 seconds keep the exchange alive.

## Generate Takes

Compose Seedance 2.5 reference takes through the gateway with both host images and both
voice references, the two-host conversation kit, the segment's dialogue projection and the
listener's silent reactions. One take per segment; the split opening is its own short take
built from a generated split image. Lifestyle B-roll is one silent multi-scene reference
clip from three lifestyle stills of the same host, or separate stills.

## Compose

Align every take (`../../skills-shared/word-timing.md`), then render with a plan
(`../../skills-shared/anchored-composition.md`): `captions` with `role_colors` per host,
the montage as a `video` layer on its selection with `mute: true`, a `split-frame` state
that outlines the active speaker if the layout stays split, a lower-third on the first
appearance, a product reveal `label` or sticker on its moment, music bed under the voice.
Check the handoffs: the montage leaves before the next line lands, the prop is in the right
hand after the cut, and the caption colour matches the speaker.

Deliver the clip and the editable project. New topics, products or hosts are variants
(`../pixverse-video-variants/SKILL.md`).

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-presenter
description: "AI presenter: make an episode led by one consistent on-screen presenter, with measured speech, supporting shots and a finished edit."
---

# Presenter Episodes

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Keep the same host recognizable across an episode while giving the explanation enough visual variety.

Read `./references/episode.md` before script splitting, prompt authoring or production.
It defines casting/bootstrap, three physical cameras, speech-derived durations, the full
voice/performance prompt, supporting-media choices, measured phrase anchors and editable
assembly. These requirements are part of a complete episode; generic calm delivery is
not an adequate substitute. A standalone avatar clip follows its narrower user brief.

## Lock The Host And Script

Use the supplied accepted image/video avatar. If no host exists, establish a concrete appearance and accept a reference before producing many dependent shots. Record face, hair, wardrobe, set and reference roles in project memory. A reference pack provides conditioning; it does not train a persistent identity model.

Write or accept the full spoken script. Divide it into coherent speaking sections with pronunciation notes. For a recurring program, preserve the host references and episode style while updating content and required factual sources.

## Plan The Performance

Choose which passages need the host on camera and which need evidence, screen capture or explanatory cutaways. Make per-shot control frames for the host's important camera positions. Avoid asking one long generation to preserve an entire episode.

For visible speech, request the accepted words in native speech shots or use an actually available supported sync service. Audio-reference conditioning is not a guarantee of identical voice. Replacing a talking clip's soundtrack with new TTS is not a lip-sync solution.

Use reaction and listening shots only where appropriate. Keep camera position, eye line and set coherent, and let natural pauses create edit points.

## Timing And Prompt Contract

Calculate each job from its actual spoken copy plus a 0.4s onset/closing allowance,
rounded up within supported limits. Chinese requires a spoken-duration estimate.
No per-cut time buffer or default long silent ending. Change physical viewpoint and
shot size at meaningful phrase transitions; do not add a cut quota. Write reference
lock, numbered physical shots, VOICE & MANNER, REALISM LAYER, exact dialogue once per
shot, semantic cut cues and a brief natural landing. The image fixes appearance, not
a frozen body pose. Keep overlay instructions in the edit plan.

## Complete The Episode

Edit accepted speaking sections and supporting visuals into the full script order. Check word accuracy, pronunciation, lip sync, voice consistency, face, wardrobe and transitions. Use captions only when requested or included in the plan. If exact speech is not achieved, identify the affected take and choose a supported repair through a new preflight when needed.

Deliver the completed episode and an episode record with host/reference IDs, accepted take IDs and reusable style choices. A series request is planned and approved in manageable episode batches.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-product-stills
description: "Product photography: create product-led studio, lifestyle or still try-on images from references while preserving shape, materials, color and branding."
---

# Product Stills

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Give each image a use: identifying the item, showing a feature, establishing scale or placing it in a believable setting.

Read `./references/shoot-recipes.md` before prompting. It defines ten modes, mode-specific
composition, three typography cases, multi-view identity chaining and targeted refinement.
Default an unspecified ordinary shoot to three distinct variants; preserve explicit count.
Do not reduce carousels or ad packs to unrelated studio images.

## Build The Shoot

1. Inspect the item before choosing a scene. Record the label, silhouette, material, moving parts and visible dimensions. Use the supplied image as a product reference, not just a mood cue.
2. Choose the requested number of compositions, or the recipe's three-variant default for
   an unspecified shoot. A carousel/ad pack needs its ordered outline before images.
3. Assign framing, surface, light and background. Reserve copy space only when requested;
   otherwise let the composition use the frame. Keep one product anchor across the set.
4. Use image reference generation through the account-compatible image route. For an invented product, establish an accepted design before producing dependent views.
5. Inspect every output for item fidelity. Finish exact labels or copy with a controlled layer when needed.

## Useful Choices

| Purpose | Art direction | What must hold |
|---|---|---|
| Catalog | Neutral background, readable contour, contact shadow | Entire item, true color, no extra packaging |
| Lifestyle | A specific plausible place and activity | Scale, real use, product remains visible |
| In-hand | A believable grip, clear fingers and supported weight | Correct grasp, hands and item proportions |
| Hero/banner | One dominant silhouette with copy space | Brand shape and crop-safe framing |
| Detail | Isolate the actual seam, connector, surface or mechanism | No invented feature |
| Static try-on | Accepted wearer and exact wearable placement | Pattern, fastening and body intersections |
| Collection | Distinct reference for each item, planned arrangement | Item count and identity |
| Concept/restyle | Change only the requested design or setting | Separate real product facts from concept changes |

For an effect-led brief, read only the matching family in `../pixverse-visual-recipes/references/image-directions.md`.

## Prompt Example

“Use the supplied moss-green insulated cup as the exact product. On a pale oak work desk, show its handle to camera-right and its closed black lid. Window light from camera-left produces a soft contact shadow. Eye-level three-quarter framing, cup occupying the left half, uncluttered space at right. Preserve the printed 350 ml mark, handle opening and matte coating. No extra straw or invented logo.”

## Deliver

Export the requested image set with purpose-based names. Compare labels, closures and color across the set. A failed angle does not invalidate the other accepted images.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-product-video
description: "Product ad: create a product-led video with identity locks, purposeful camera work, readable benefits and a final edit."
---

# Product Videos

For “Adapt this ad using my product”, use reference information already obtained by Codex or
provided by the user: the desired mood, pacing, camera choices, shot notes or usable assets.
Codex chooses how to acquire and understand a linked reference with its available tools.
Continue here with an original product brief, storyboard and production plan; no plugin-owned
video analysis is a prerequisite. Use known product facts and distinguish any creative
assumptions from information observed in the reference.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Make the product the star. Preserve packaging, silhouette, material, logo placement, and reason to care.

Read:

- `../../skills-shared/prompt-craft.md`
- `../../skills-shared/model-routing.md`
- `../../skills-shared/capability-priority.md`
- `../../skills-shared/cinematography.md`

## Fast Path

1. Create or read project memory.
2. Classify reference: product identity, brand style, or mood only.
3. Choose one clear direction when the brief is concrete. If direction is open, propose short textual options; generate only requested/approved candidates. Add a control image only when product accuracy or continuity needs it. For effect direction, see `../pixverse-visual-recipes/SKILL.md`.
4. Render the requested duration and count through the account-compatible route. Multi-shot work uses accepted product references and individual shot boards. Standalone still deliverables use `../pixverse-product-stills/SKILL.md`.
5. QA for product fidelity, crop, logo/text drift, and useful motion.

Use `seedance-2.5` at 1080p for both preview and final shots. Apply `../../skills-internal/pixverse-seedance-prompt-enhance/SKILL.md` before preflight; preserve the workflow-specific creative controls.

If a premium image-model comparison is requested, compare GPT Image 2.5 Sunburst High and Nano Banana Pro
(`gemini-3.0`) on the same brief. Include both tasks in preflight; comparison is optional.

## Creative Variants

- **Quiet luxury:** macro material, warm side light, slow gestures, minimal copy.
- **Demo clarity:** hand/use-case, before-after, benefit made visible.
- **Thumb-stopper:** odd scale, surreal environment, striking first frame.
- **Brand film:** human emotion around product, less literal demo.
- **Launch teaser:** abstract reveal, final locked packshot.

## Prompt Anchors

Always include:

- product category and exact visual identity
- material and scale
- what must not change
- camera rhythm
- lighting
- final beat

For text/logo, prefer a locked still image plate. Video models are weak at precise readable text.

## Queue Shape

Image board:

```bash
pixverse create image --model gpt-image-2.5-sunburst --quality 1440p --detail-level high --aspect-ratio 9:16 --prompt "..."
```

Video:

```bash
pixverse create reference --model seedance-2.5 --quality 1080p --duration 15 --aspect-ratio 9:16 --images {{board.path}} --prompt "..."
```

Use `pvx queue` for board-to-video chains.

## Make It Yours

If the user says "cool" or "not generic", do not just add neon. Change the creative mechanism: unusual reveal, product-as-environment, tactile macro detail, social-native imperfection, or a symbolic human moment.

## Quality Checks

- product still recognizable
- logo/label not hallucinated beyond tolerance
- no fake UI or fake claims
- first frame works as thumbnail
- final frame can act as product packshot

Referenced files: 1

View saved version →

---
name: pixverse-ranking-video
description: "Ranking video: a host argues a tier list, top-N or comparison while an on-screen board keeps every verdict, with word-anchored reveals, karaoke captions, B-roll jokes and reveal sounds."
---

# Ranking Video

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

A ranking makes an argument through comparison. The board remembers earlier judgments while
the host, evidence and performance explain the next one. The picture is generated; the board,
captions, inserts and sounds are composed locally and land on the words that decide them.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Shape The Argument

Resolve the subjects, the verdict for each, the board type (tier rows, numbered column, top
three), the host and the tone. Preset entries can already occupy the board when the clip
opens; the spoken passages introduce the subjects being judged now. Each subject gets a
verdict line with a clear verdict word, a reason with comic or factual evidence, and an
optional insert that illustrates the joke or claim.

Write the script with `../../skills-shared/semantic-script.md`: a moment on each verdict word
(`@{tea-verdict!}D tier`), a selection around each explanation that an insert should cover,
cue breaks by reading phrase. Measure it; ranking lines read fast, so choose take durations
with a small margin.

## Produce The Material

- Host image: `../../skills-shared/prompt-kits.md` phone-capture kit, host toward one side,
  the board's region calm. Reuse a supplied person as the identity reference.
- Speaking take(s): `../pixverse-talking-head/SKILL.md` with the script's dialogue, one take
  per segment; keep attitude and eye contact in the prompt.
- Subject icons: supplied photographs or logos for real subjects (identifiers are never
  generated lookalikes); generated stills for fictional or generic subjects.
- Inserts: one Sunburst still per joke or claim with its own visual language; no board
  furniture in the picture.
- Sounds: builtin `pop`/`ding` reveal sounds are enough; add music through `../pixverse-audio/SKILL.md`
  when the brief wants a bed.

## Compose

Follow `../../skills-shared/word-timing.md` to align each take, then render one board state
per verdict with the `tier-board` or `ranked-column` component
(`../../skills-shared/graphics-components.md`) and bind states, inserts, captions, verdict
stamp and sounds in a plan (`../../skills-shared/anchored-composition.md`). The reference
plan in `./references/board-recipe.md` is the starting point.

## Check And Deliver

Sample the frames on each verdict word and each insert handoff. The board must change on
its word, remain readable at delivery size and never cover the host's face; inserts must
leave before the next claim. Deliver the video with the script, timeline, plan and graphics
so a new subject list or a new host is a re-run, not a rebuild
(`../pixverse-video-variants/SKILL.md`).

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-setup
description: "PixVerse setup: repair CLI or FFmpeg dependencies, authenticate and diagnose account readiness when setup blocks the requested workflow."
---

# PixVerse Setup

Use this when the user asks to install, authenticate, diagnose, check balance, inspect slots, fix PixVerse CLI problems, or understand command/queue syntax.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

Read the internal playbooks that match the issue:

- first-run setup, dependencies, login, FFmpeg, doctor: `../../skills-internal/pixverse-agent-setup/SKILL.md`
- command forms, model ids, queue specs, task polling, CLI errors: `../../skills-internal/pixverse-agent-gateway/SKILL.md`
- Canvas binding, graph sync/mutation, paid reconciliation, browser handoff: `../../skills-internal/pixverse-agent-canvas/SKILL.md`
- project bootstrap after setup is ready: `../../skills-internal/pixverse-agent-project-memory/SKILL.md`

Read `../../skills-shared/web-handoff.md` before any OAuth, subscription, workspace-management, or
Canvas browser action. Every plugin web destination uses the Codex in-app Browser by default.

The latest reviewed npm release is `1.4.5` (reviewed 2026-09-22). `1.4.0` is the compatibility floor, not an install pin. Refresh the online runtime with `"${PVX}" bootstrap --yes` when the user asks to update or when installing/replacing this plugin. A global CLI upgrade does not refresh this private runtime. Read `../../skills-shared/pixverse-cli-1.4.5.md` for the current CLI changes.

## Setup Gate

PixVerse Agent Plugin `1.3.2` requires `pixverse>=1.4.0` and Node.js `>=22.12.0` on both managed channels, and never uses a global `pixverse` package. Setup reports `pixverse_cli_channel`: the online plugin installs npm `latest` in its private stable runtime, while the local plugin installs only the internal ZIP bundled with that plugin into a separate internal runtime. A local package is rejected if its bundled CLI declares a different Node engine.

Start with cached readiness:

```bash
"${PVX}" setup status
```

If status is missing or blocked, refresh checks:

```bash
"${PVX}" doctor
```

If doctor reports the managed CLI missing or below minimum, or when installing/updating the plugin, refresh it in one call:

```bash
"${PVX}" bootstrap --yes
```

On the online channel, this installs `pixverse@latest` when absent and refreshes an existing supported version. On the internal channel, it verifies, extracts, dependency-installs, smoke-tests, and activates the bundled internal ZIP. A matching internal runtime is reused only after its install metadata, package/capabilities, dependencies, entrypoints, and smoke test still pass; otherwise bootstrap rebuilds it from the ZIP. Internal installation never falls back to npm latest, and a failed update preserves the previously active internal CLI.

Doctor probe timeouts are returned as structured blocked checks and cached with next steps; they must not surface as Python tracebacks or be mistaken for missing binaries. Retry doctor after checking connectivity, and only start paid work once auth and slot capacity are readable.

Upgrading the plugin/CLI (including replacing a same-version package) must preserve existing CLI
credentials and reuse the Codex IAB profile. Do not run login/logout or clear authorization solely
because the version changed; an expired or revoked session still follows the normal login flow.

If PixVerse auth is missing:

```bash
"${PVX}" pixverse auth login --json
"${PVX}" doctor
```

Run login as one long-lived call. The wrapper forces JSON mode, preventing PixVerse CLI from opening
the system browser. As soon as stderr emits `Authorize at: <url>`, open that exact URL in the Codex
in-app Browser and keep the login process alive while the user authorizes. Never ask them to paste
account tokens. After completion, `doctor` refreshes cached auth/readiness. If the URL is not emitted or
login remains false, run `"${PVX}" pixverse auth status --json`, check connectivity, and retry login once
rather than attempting paid work. Use `--open-system` only after an explicit user request.

Use the wrapper for realpath-safe CLI calls:

```bash
"${PVX}" pixverse --version
"${PVX}" pixverse auth status --json
```

## Canvas Setup Routing

Canvas setup, login, account, balance, membership, and active-workspace failures still use the setup
gates in this skill. For project binding, graph checkpoints, guarded mutations, paid node generation,
credit reconciliation, and browser handoff, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md`; that is the canonical Canvas playbook.
Canvas Browser login and PixVerse CLI authentication are separate sessions.

## Balance And Slots

Before paid generation or queue planning:

```bash
"${PVX}" billing snapshot
"${PVX}" quote queue <queue.json>
```

After setup, generation proceeds automatically after preflight by default, including the first batch. Only effective `require` waits for batch approval. If the user asks to control spending, enable project confirmation; “Allow future generation” restores automatic execution. Follow `../../skills-shared/generation-confirmation.md`.

The same preflight reports login and membership. Follow `../../skills-shared/quality-policy.md`:
Free/Basic must stop and choose upgrade or fallback; any model-entitlement rejection also
pauses generation. Show the clickable subscription/recharge link, then wait for the user's
choice. Upgrade is followed by a live recheck; fallback consent enables v6 540p / Nano Banana
2 Lite 1080p with a fresh preflight. Obtain the subscription URL using:

```bash
"${PVX}" pixverse subscribe
```

Do not call `subscribe` without the user's intent. It returns a structured handoff that must be opened
in the Codex in-app Browser; it does not open the system browser by default. If a user explicitly
confirms that their displayed Free/Basic account is an unrestricted internal test account, record the
local exception with `"${PVX}" preferences membership-routing unrestricted-test`; never infer or offer
that override to ordinary users. It changes model routing only—login, balance, quote, and approval
gates remain active.

Referenced files: 1

View saved version →

---
name: pixverse-short-drama
description: "Short drama: develop scripted scenes with recurring characters, dialogue coverage, continuity and separately scoped generation batches."
---

# Short Drama

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

Vertical drama is about emotional compression and a clear next-episode hook.

Read:

- `../../skills-shared/prompt-craft.md`
- `../../skills-shared/audio-craft.md`
- `../../skills-shared/model-routing.md`

## Fast Path

1. Pick one conflict and one reversal.
2. Lock 1-3 characters.
3. Write a 4-beat vertical episode:
   - hook
   - confrontation
   - reveal
   - cliffhanger
4. Use 9:16, 8-15s.
5. Use dialogue only if it can be short and readable.

## Genres

- revenge reveal
- CEO romance power flip
- family betrayal
- rebirth/second chance
- hidden identity
- contract marriage
- workplace humiliation reversal

## Character Rules

If the user wants a series, use `../pixverse-character-sheet/SKILL.md` first or create character boards. Consistency beats novelty.

## CLI Shape

```bash
pixverse create video --model seedance-2.5 --quality 1080p --duration 15 --aspect-ratio 9:16 --prompt "..."
```

Use `seedance-2.5` at 1080p for both preview and final shots. Apply `../../skills-internal/pixverse-seedance-prompt-enhance/SKILL.md` before preflight; preserve the workflow-specific creative controls.

For exact spoken lines, consider standalone voice plus post-editing.

## Quality Checks

- hook visible in first 2 seconds
- characters readable on mobile
- reversal lands
- cliffhanger is visual, not just explained
- no dialogue overload

Referenced files: 1

View saved version →

---
name: pixverse-street-interview
description: "Street interview: a question-and-answer encounter on location with interviewer, guest and shared views, timed reveals such as an answer strip or emoji board, reaction cuts and readable captions per speaker."
---

# Street Interview

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

A motivated encounter: someone stops someone, asks a question, and the answers arrive as
reveals the audience collects. The picture alternates interviewer, guest and shared views;
an answer strip or emoji board fills in on the spoken words; captions identify speakers.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Set Up The Encounter

Resolve the location, the interviewer (often off-camera or a mic-holding hand), the guest,
the question and the list of answers. Give the encounter a reason: something the guest has
or did that makes the question worth asking. Generate the guest's location portrait with the
phone-capture kit (`../../skills-shared/prompt-kits.md`), then complementary views: a
two-shot over the interviewer's shoulder, a tighter guest view for answers, an environment
view that shows the motivating detail. Real places and people come from supplied material.

## Script The Reveals

Use `../../skills-shared/semantic-script.md`: `ASKER:` and `GUEST:` turns, a moment on each
answer word (`@{rule-one!}Manifest.`), selections around the guest's explanations for
reaction cuts, cue breaks that isolate the answer words so captions land them alone. Measure
segments; keep each answer in its own 5–8 second take so a retake never touches accepted ones.

## Generate

Seedance 2.5 reference takes through the gateway with the guest image, the two-shot image
for shared views and a voice reference per speaker. Direct the physical exchange: the mic
enters from the interviewer's side, the guest looks slightly off-lens toward the asker,
reactions are small and quick. Handheld energy is welcome; keep the guest's framing region
stable so the strip and captions have a home.

## Compose

Align (`../../skills-shared/word-timing.md`) and render with a plan
(`../../skills-shared/anchored-composition.md`): a `reveal-strip` component whose states
reveal one slot per answer moment, `ding` or `pop` sounds on the same moments, captions with
`role_colors` per speaker, an environment still cut in on the motivating detail, a final
`label` summarizing the answers if the format wants it. Read `./references/reveal-board.md`
for the strip and reaction timing.

Check that each slot reveals on its word, reaction cuts return before the next answer, and
captions never cover the strip. Deliver the clip with script, timeline, plan and graphics;
new guests, languages or answer sets are variants (`../pixverse-video-variants/SKILL.md`).

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-studio
description: "Workflow discovery: choose a PixVerse capability or shape an underspecified brief. Concrete media, editing and Canvas tasks use their direct workflows."
---

# PixVerse Studio

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Help the user choose the existing workflow that fits the requested result.
This is a planning and discovery entry, not a prerequisite for other public skills.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## First Move

Canvas first-confirmation fast path:

1. If the user asks to create, open, continue, inspect, edit, sync, or generate in a PixVerse Canvas
   project or shared Agent/Web editor, read `../../skills-internal/pixverse-agent-canvas/SKILL.md` and
   execute its early browser handoff before ordinary preparation. Merely starting a new Codex task is
   not Canvas intent and does not open a browser automatically. This route short-circuits generic
   production routing until the first useful paid confirmation: keep only this Studio entrypoint, the
   Canvas playbook, and `web-handoff.md` on that critical path. Do not also expand production,
   gateway, project-memory, creative-orchestration, or their references unless an unresolved blocker
   or a field needed by that first batch cannot be handled by the Canvas playbook itself. Load those
   playbooks after confirmation or at the later stage that actually needs them.
   Continue to use the Canvas playbook's **Cloud Preview And On-Demand Delivery** contract after
   confirmation. It overrides this entrypoint's queue/local defaults: preview in Canvas, reuse cloud
   references, and do not automatically download, run local QA, or query post-generation credits.
   A directory binding alone never activates that policy. For Canvas continuation use
   `project resume [hint] --surface canvas --format markdown`; use the same explicit surface on
   portfolio and handoff. No local media is required for cloud delivery.
2. For ordinary generation, select the matching public workflow below and read it directly.
   Read `../../skills-shared/cli-workflow.md` when execution starts; it preserves the existing
   preflight, paid retry, preview and handoff rules.
3. For setup, login or dependency problems, use `../pixverse-setup/SKILL.md`.
4. For project recovery, inspection or file delivery, use `../pixverse-delivery/SKILL.md`.

## Existing Workflows

| Requested result | Public workflow |
|---|---|
| Create Images | `../pixverse-create-image/SKILL.md` |
| Create Videos | `../pixverse-create-video/SKILL.md` |
| Voice and Music | `../pixverse-audio/SKILL.md` |
| Product Videos | `../pixverse-product-video/SKILL.md` |
| UGC Videos | `../pixverse-ugc-video/SKILL.md` |
| Music Videos | `../pixverse-music-video/SKILL.md` |
| Cinematic Story Clips | `../pixverse-cinematic-story/SKILL.md` |
| Short Drama | `../pixverse-short-drama/SKILL.md` |
| Game Trailers | `../pixverse-game-trailer/SKILL.md` |
| Character Sheets | `../pixverse-character-sheet/SKILL.md` |
| Video Editing | `../pixverse-video-editing/SKILL.md` |
| Product photography and still try-on | `../pixverse-product-stills/SKILL.md` |
| Product detail cards and exact specifications | `../pixverse-listing-images/SKILL.md` |
| Video thumbnails and covers | `../pixverse-cover-art/SKILL.md` |
| Specific image treatments or short motion effects | `../pixverse-visual-recipes/SKILL.md` |
| Try-on and wearable demonstrations | `../pixverse-fashion-video/SKILL.md` |
| Package opening and product reveals | `../pixverse-unbox-video/SKILL.md` |
| Step-by-step product operation | `../pixverse-howto-video/SKILL.md` |
| Independent variations of a supplied ad | `../pixverse-ad-variants/SKILL.md` |
| Complete narrator-led explainers | `../pixverse-explainer/SKILL.md` |
| Recurring presenter episodes | `../pixverse-presenter/SKILL.md` |
| Full spoken scripts without media generation | `../pixverse-video-script/SKILL.md` |
| Natural narration or fixed-window takes | `../pixverse-voiceover/SKILL.md` |
| Captions burned into video | `../pixverse-captions/SKILL.md` |
| Fast music-led motion, beat-synced typography and graphics | `../pixverse-motion-design/SKILL.md` |
| Generated brand visuals and mockups | `../pixverse-brand-system/SKILL.md` |
| Rebuild a reference video as your own version | `../pixverse-video-remake/SKILL.md` |
| Change the look of existing footage, keep its motion | `../pixverse-video-restyle/SKILL.md` |
| Tier lists, top-N, comparisons with a live board | `../pixverse-ranking-video/SKILL.md` |
| Two-host podcast or interview clips | `../pixverse-podcast-clip/SKILL.md` |
| Street interviews with timed reveals | `../pixverse-street-interview/SKILL.md` |
| Speaking takes with word timing | `../pixverse-talking-head/SKILL.md` |
| Many versions of one generated project | `../pixverse-video-variants/SKILL.md` |
| The same video in another language | `../pixverse-video-translate/SKILL.md` |

For unresolved production depth, model-route comparison, or production scaffolds, read
`../../skills-internal/pixverse-production/SKILL.md` only when that guidance is needed.
Read `../../skills-shared/web-handoff.md` before OAuth, subscription, workspace-management
or Canvas browser work.

Existing projects can be recovered in one call:

```bash
"${PVX}" project resume [title-or-slug-hint] --format markdown
```

Use `project portfolio` only if several projects may match. For Canvas resume/portfolio/handoff,
select `--surface canvas`; its **Cloud Preview And On-Demand Delivery** contract overrides
ordinary queue/local defaults. A directory binding alone does not activate Canvas.

Prompt enhancement also runs automatically before Seedance 2.5 generation.

Referenced files: 1

View saved version →

---
name: pixverse-talking-head
description: "Talking head: one person speaks a script straight to camera as reusable A-roll, with a locked identity, a consistent voice, measured durations and word timing delivered for captions and graphics."
---

# Talking Head

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

Produce the spoken performance itself: a person delivers exact words to camera with a stable
identity and voice, in takes sized to the script. The output is A-roll plus measured word
times, ready for captions, boards, inserts and stickers in other workflows, or delivered as
is. Creator ads with products use `../pixverse-ugc-video/SKILL.md`; full hosted episodes use
`../pixverse-presenter/SKILL.md`.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Lock Words, Person, Voice

Write or accept the script in `../../skills-shared/semantic-script.md` form: segments per
take, cue breaks, pronunciation for names and numbers. Measure each segment
(`"${PVX}" script measure ... --rounding ceil`) and pick a supported duration with a small
margin; natural takes add pauses.

Identity: a supplied portrait or an accepted generated image from the phone-capture kit
(`../../skills-shared/prompt-kits.md`), seated or standing where the brief needs, with calm
space where graphics will go. Voice: one short TTS reference line (`create voice`, a preset
that fits the person) reused on every take as an `--audios` reference. Record image, voice
and wardrobe in project memory so later segments and variants keep them.

## Generate Takes

Seedance 2.5 `create reference` per segment through the gateway: host image, voice
reference, the speaking-take kit with the segment's `dialogue` projection as the exact
words, one physical beat per phrase, composition stability and camera axis chosen for the
format. No captions, on-screen text or music in the take. Submit segments as one queue;
each is independent, so a retake replaces only its own segment.

## Measure And Deliver

Build the timeline (`../../skills-shared/word-timing.md`) as soon as takes download. Check
alignment confidence, unmatched words and extra ASR words. Listen to flagged passages:
correct the transcript if recognition is wrong; only a verified skipped or added spoken
word calls for a retake of that segment under the effective generation policy. Deliver the takes, the timeline JSON and, when asked, captions
(`"${PVX}" timeline captions ... --style clean|karaoke`) or a finished composition through
`../../skills-shared/anchored-composition.md`. Read `./references/performance.md` for
delivery direction and multi-take continuity.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-ugc-video
description: "UGC ad: create a creator-led or product-led social ad, with a clear hook, believable demonstration and measured spoken delivery."
---

# UGC Videos

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

For ordinary creation or editing, read `../../skills-shared/cli-workflow.md` before
executing this workflow. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

## Select And Read The Full Recipe

| Intent | Required reading | Production structure |
| --- | --- | --- |
| Product hero, casual UGC, off-screen voice; no speaking face | `./references/product-demo.md` | Four panels/cuts: intro, two distinct demonstrations, result |
| Visible creator review, topic, routine or story | `./references/creator-story.md` | Eight panels/cuts; story arc, native performed speech, optional product |
| Explicit silent single product beat | `../pixverse-create-video/SKILL.md` or `../pixverse-product-video/SKILL.md` | Follow that direct brief |

Read the selected reference and `../../skills-shared/ugc-production.md` before writing
boards or prompts. Their cut counts, time budgets, mandatory board refinement, reference
order, hand/prop rules and speech contract are required parts of these named recipes.
Generic short-video simplicity does not remove them. Explicit user shot lists and tone
choices override defaults. Preserve no-face intent and accepted identities.

For product mode, use four distinct physical events and native off-screen VO; auxiliary
people remain silent. For creator mode, a productless topic/routine/story is a complete
supported mode: omit product references instead of inventing a brand or sales prop.
An open brief needs a story, not eight repeated product poses. Do not invent purchases,
customer testimonials, clinical outcomes or unsupported measurable benefits.

Split longer work into 4-15s clips using the shared duration rule. Build and refine boards
sequentially, then write all motion prompts with full numbered cuts and exact dialogue.
No default extra hook variants, music, external TTS or additional posts. Series intent is
different from one multi-board video.

## Related Deliverables

- Wearing or fit: `../pixverse-fashion-video/SKILL.md`
- Package opening and reveal: `../pixverse-unbox-video/SKILL.md`
- A repeatable physical procedure: `../pixverse-howto-video/SKILL.md`
- Independent edits of an existing ad: `../pixverse-ad-variants/SKILL.md`
- Deliberately art-directed brand film: `../pixverse-product-video/SKILL.md`

## Finish And Revise

Watch the complete action. Check that the visible proof supports the spoken claim, the item remains the same, contact is plausible and dialogue is intelligible. Listen to native audio before replacing it. Deliver the completed clip with requested exact overlays. Preserve accepted shots; fix only the failed action, spoken take or local layout.

For prompt-only work, deliver the complete board plan, refinement prompt, all four or eight timed motion sections, exact speech and reference roles without account or generation calls. Do not describe planned control assets as generated.

Referenced files: 3

View saved version →

---
name: pixverse-unbox-video
description: "Unboxing video: build a believable package-opening sequence whose reveal leads to the actual product, with consistent hands, packaging and sound."
---

# Unboxing Video

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Make the reveal follow the package's real construction. Read `./references/production.md`
before the board or prompt: the four-part arc, opening branches, timing, performance,
hand allocation and closing transition are core production rules.

## Establish The States

Inspect the outer package, opening mechanism, inner tray and verified contents. The standard
arc begins with a closed package. If only an open reference exists, reconstruct a closed
state only when the visible parts support it; otherwise resolve missing evidence. Preserve
an explicit already-open brief, but do not silently choose it for a requested full unboxing.

Use four active 9:16 slots on a 21:9 board: PACKED → REVEAL → PRODUCT-FOCUS → SATISFACTION.
Create and refine the board, then use clean board + creator + product for the video.
One package in cut 1; its correct opening ends that cut. The product emerges in cut 2;
the package may remain at an edge there, then disappears for cuts 3–4. One product and
two explicitly assigned hands; don't re-open or re-pack between beats.

Use a visible creator by default, preserving an explicit hands-only request. Default POV is
fixed → fixed → fixed close-up → selfie. For 12s, allocate 3 / 3.5 / 2.5 / 3 seconds.
Write four timed Cut sections. Usually there are three hard cuts; on an eligible closing
clip the reference requires one continuous Pick-Up instead of the final boundary.

## Write The Performance

Each cut describes one distinct action in motion, real grip/weight, at least five small
observable changes and its exact dialogue/sound. First sound within .4s; natural anticipation,
one real reveal peak, detail observation, satisfied resolution. Keep mouths unobstructed
for visible speech. No slow-lift/slow-rotate/long-hold default. The reference supplies
the package-specific mechanics and caption/ending rules; do not substitute a generic box.

If a long opening exceeds a credible short shot, divide it at a real state change. Do not cover impossible package motion with a rapid cut and claim the demonstration was checked.

## Finish

Check hands, lid geometry, contents count, insert and the reveal's clarity. Preserve native package sounds when useful. For a reaction line, keep spoken experience fictional or user-supplied instead of asserting a real testimonial. Add captions only when requested. Deliver the completed edit; repair only the broken state or transition.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-video-editing
description: "Video editing: edit existing footage from any source; trim, reorder, reframe, retime and mix sound into a finished video. Local edits need no PixVerse login."
---

# Video Editing

Edit the supplied footage directly, including clips made outside PixVerse. Keep source files
intact and write a new master under `projects/<slug>/deliverables/`. A request to cut, reframe
or mix existing media needs no PixVerse login, model lookup, generation queue or paid preflight.
Use FFmpeg/ffprobe for local edits and preserve the supplied source media.

For layered layouts, animated overlays or revision of an existing FFmpeg edit, read
`./references/composition-contract.md`. It preserves source intervals, timing, audio
ownership and reproducible render files. Keep unrelated existing edits intact.

## Fast Path For Existing Footage

Work from usable footage already supplied or obtained by Codex. If the user supplies a
webpage link, Codex chooses how to acquire the material with its available tools before
editing here. Reuse any reference notes that are already available.

1. Inspect available files and needed stream metadata in one local call. Reuse known duration,
   dimensions and sound decisions. View/listen only where the requested edit needs judgment.
2. Translate the brief into source intervals, order, output aspect and sound treatment. For a
   simple trim, the request itself is the edit plan. For several shots, save a compact cut list
   with source path, in/out seconds, output position, overlays and audio treatment.
3. Render with FFmpeg. Use `./references/local-editing.md` for accurate trims,
   mixed-format assembly, reframing and speed changes. Batch compatible operations into one
   render rather than repeatedly encoding a clip.
4. Return the actual output file and the cut list or render source. Verify the requested change:
   for example, duration for a trim or subject framing for a vertical crop. Add other checks
   only when the edit requires them or the user requests review.

For reference-based editing, reuse timecoded notes already supplied by the user or Codex;
there is no prerequisite PixVerse analysis workflow. For captions use `../pixverse-captions/SKILL.md`; for exact animated layers use
`../pixverse-motion-design/SKILL.md`; for inserts, boards, stickers and sounds that must land on
spoken words, use `../../skills-shared/anchored-composition.md` (`"${PVX}" timeline render`). Read `../../skills-shared/audio-craft.md` for a sound rebuild,
not for every trim. Only missing generated media enters the CLI contract below.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## CLI Execution

Apply `../../skills-shared/quality-policy.md` to all generated images/videos: Sunburst 2K/high,
Seedance 2.5 1080p with automatic prompt enhancement, and an explicit upgrade/fallback
choice for Free/Basic or model-entitlement rejection.

For an explicitly selected Canvas project or node, read
`../../skills-internal/pixverse-agent-canvas/SKILL.md` first and follow its Canvas
preflight and delivery contract; the ordinary queue/local route below does not replace it.

When the edit requires new generated media, read `../../skills-shared/cli-workflow.md` before
that generation. It contains the existing account-aware routing, paid-work
confirmation, progress, preview and project-handoff rules. Studio and Production are
not prerequisites. The shared account, confirmation and direct-medium rules govern the
examples and defaults below. Read the gateway only when writing manual CLI commands or queue specs.
Creation command examples below are queue-task fragments, not permission to submit paid
`pixverse create` commands directly.

## Canvas Boundary

Apply this boundary only after the current target has been resolved as Canvas under
`../../skills-internal/pixverse-agent-canvas/SKILL.md`. For that Canvas target, inspect the current Canvas capabilities and
node schemas before downloading assets or invoking FFmpeg, then use Canvas-native editing nodes and
existing graph connections whenever they can express the requested composition.

For every non-Canvas project—including ordinary PixVerse queues, standalone local assets, and local
deliverables—do not run Canvas discovery or change the established local editing route. The recipes
below remain the default and behave as before. They also remain available as an explicit Canvas
fallback; they do not override the Canvas-native-first rule once Canvas is the selected target.

Do not download, locally edit, and re-upload media that already exists in Canvas merely because a
local recipe is available. If Canvas lacks a required operation or the user explicitly requests a local
master, keep intermediates local and upload only the final result once if later Canvas work actually
needs it.

## Common Tasks

- download PixVerse assets
- stitch multiple clips
- add background music
- overlay narration
- strip rejected generated audio
- rebuild sound from separate voice, music, ambience, and SFX stems
- burn subtitles
- trim starts/ends
- create final deliverable folder
- verify the requested edit

## Subtitle And TTS Sync

For narrated videos, use actual audio as the clock and keep one canonical timed subtitle file for rendering. An SRT entry looks like:

```text
1
00:00:00,000 --> 00:00:03,200
First spoken subtitle line
```

Workflow:

- Reuse accepted audio, or produce narration in natural sections through `../pixverse-voiceover/SKILL.md`. Write `projects/<slug>/prompts/narration.srt` from verified words and actual timing through `../pixverse-captions/SKILL.md`.
- Keep entries readable in one or two lines with sensible phrase boundaries. Match the requested typography; clean display punctuation only when that style is desired, retaining the spoken source text.
- Inspect the timed captions: `"${PVX}" subtitles inspect projects/<slug>/prompts/narration.srt`.
- For an explicit fixed-window take list, `"${PVX}" subtitles voice-queue projects/<slug>/prompts/narration.srt projects/<slug>/voice-queue.json --project <slug> --segments-dir projects/<slug>/prompts/tts-segments --voice-id <voice_id> --language zh` can prepare separate takes. It is not the default for every captioned film; preserve the voice model's natural speed unless the user requests a speed change.
- Quote and confirm a new voice queue before running it. Do not create voice tasks for a caption-only change.
- Measure each returned take. Align the final subtitle windows to speech, or fit explicitly timed takes to their required windows without clipping words. One caption does not need to own one voice asset.
- Burn subtitles from the same SRT, or convert the same SRT to ASS for style. Use the plugin default subtitle style unless the user asks for a different look.
- If speech and picture lengths differ, align captions to real speech and trim only verified
  idle padding. Fixed-window narration follows its voiceover recipe; do not tempo-shift or
  cut words merely to pass its timing gate. Explicit user-requested retiming remains available.

Example burn-in:

This example explicitly replaces the soundtrack with accepted narration. For a caption-only
change, preserve the existing audio and use the captions workflow instead.

```bash
"${PVX}" subtitles style
"${PVX}" subtitles style --format force-style
ffmpeg -nostdin -n -i video.mp4 -i voice.mp3 -vf "subtitles=projects/<slug>/prompts/narration.srt:force_style='<force_style_from_pvx>'" -map 0:v -map 1:a -c:v libx264 -c:a aac out.mp4
```

Default burn-in style for 1080p horizontal delivery is intentionally restrained: `Fontsize=18`, `MarginV=28`, bottom-centered, white text, small dark outline. Do not improvise oversized subtitle text such as `Fontsize=34` or larger for normal subtitles; reserve larger type for title cards or intentional on-screen typography.

## Download

```bash
"${PVX}" pixverse asset download <id> --type video --dest ./projects/<slug>/assets/videos
"${PVX}" pixverse asset download <id> --type audio --dest ./projects/<slug>/assets/audio
```

## Stitch With FFmpeg

For an ordered non-Canvas PixVerse story run, prefer the integrated local path:

```bash
"${PVX}" story assemble <slug> --output projects/<slug>/deliverables/final-story.mp4 --sample-frames
```

It resolves `shot-*` assets from the latest billing ledger, downloads missing clips, re-encodes a compatible H.264/AAC timeline, runs duration QA, samples the full timeline, and records the result. Use the manual FFmpeg forms below for custom clip order, mixed formats, transitions, or sound rebuilds.

For streams with matching codec, dimensions, frame rate, time base and audio layout,
create a concat list (escape quote characters in actual paths):

```text
file 'clip1.mp4'
file 'clip2.mp4'
```

Then:

```bash
ffmpeg -nostdin -n -f concat -safe 0 -i list.txt -c copy ./projects/<slug>/deliverables/final.mp4
```

If streams differ, normalize each clip before concatenating; see
`./references/local-editing.md`. Merely changing the output codec after a concat demuxer
does not normalize incompatible source dimensions or stream layouts.

## Add BGM

Keep music lower than dialogue. For a clip with an existing audio stream, this simple
overlay keeps that soundtrack and limits the mix to its duration:

```bash
ffmpeg -nostdin -n -i video.mp4 -i music.mp3 -filter_complex "[1:a]volume=0.18[a1];[0:a][a1]amix=inputs=2:duration=first:dropout_transition=2[a]" -map 0:v -map "[a]" -c:v copy -c:a aac out.mp4
```

When the source has no audio, map the music directly and bound it to the video duration.
Choose trim, fade, silence or a loop deliberately when the music is shorter; do not reference
a nonexistent `[0:a]` or let `-shortest` accidentally cut the picture.

## Rebuild Rejected Audio

If generated video audio contains unwanted music, bad ambience, or unusable SFX, strip it first:

```bash
ffmpeg -nostdin -n -i video.mp4 -map 0:v -c:v copy -an picture-clean.mp4
```

For serious sound work, prepare `development/sound-cue-sheet.md` before mixing:

| Time | Visual beat | Stem | Source file | Mix note |
|---|---|---|---|---|
| 00:00-00:02 | door opens | hinge creak | assets/sfx/hinge.wav | quiet, left |
| 00:02-00:06 | town exterior | ambience | assets/sfx/night-town.wav | low bed |

Do not bury a rejected generated audio track under BGM. Replace it with controlled stems or deliver a clean temp mix and state which stems are still missing.

## Targeted Verification

Inspect the requested edit, not a compulsory full production checklist. Duration, dimensions
and audio-stream presence can be checked with ffprobe; cut continuity, legibility and audible
mixing require viewing or listening when those operations were requested. For explicit QA or
a named specialized check, read `../../skills-internal/pixverse-agent-quality/SKILL.md` and use:

```bash
"${PVX}" qa inspect ./projects/<slug>/deliverables/final.mp4 --project <slug> --sample-frames
```

## Checks By Edit Type

- Cut/assembly: requested intervals and order, expected duration, no unintended black gaps.
- Reframing: requested dimensions, correct pixel aspect, important subjects still visible.
- Sound: retained/replaced/removed as requested; listen if mixing or synchronization changed.
- Captions: measured speech timing, readable type, safe placement and requested typography.
- Handoff: actual master path; include editable sources when requested and record multi-stage
  project decisions for continuation. Name the FFmpeg source list, timing data and filter graph.

Referenced files: 3

View saved version →

---
name: pixverse-video-remake
description: "Reference remake: rebuild a supplied or linked video as your own version with a new person, product, wording, language or platform, keeping the structure, timing relationships and graphics that made the original work."
---

# Video Remake

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

Start from the reference the user supplied (a file or a link) and the change they want:
their face, their product, a new audience, another language, another aspect. The result is
an original, editable production that keeps the reference's working relationships, not a copy
of its pixels.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Read The Reference

Follow `../../skills-shared/reference-breakdown.md`: get the file (`media fetch` for a link
when `yt-dlp` is installed, otherwise ask for the file), probe it, transcribe its speech,
make word-labelled contact sheets, and write `ANALYSIS.md` and `TIMELINE.md` under
`projects/<slug>/reference/`. Explain what the piece does for the viewer and which visual
and sound systems carry it: host performance, captions, board or list, B-roll, stickers,
titles, music. Name what each event responds to (a word, a claim, a reveal), because those
relationships are what the remake preserves when the seconds change.

Share the reading before generating: a short whole-piece explanation, the systems you will
keep, adapt or replace, and open questions about the user's material. Routine progress
needs no approval; a material creative choice does.

## Transform The Roles

Write the target `Brief` (what the user wants) and `Treatment` (how the new piece works) in
the project notes. Then apply `./references/transformation-lenses.md`: change the person,
product, wording, language, platform or style across every place that role appears. A
product is a prop, a screen, a claim and the closing call to action; a presenter is also the
voice, the lifestyle B-roll subject and the face inside a sticker. Preserve the purpose of an
action while giving it a form that fits the new body, product and platform.

## Rebuild The Structure

1. Write the new script with `../../skills-shared/semantic-script.md`: cue breaks where the
   reference's captions breathed, selections where its B-roll covered explanations, moments
   where its board or stickers landed. Measure each segment before choosing take durations.
2. Generate material with the matching workflows: host image and speaking takes through
   `../pixverse-talking-head/SKILL.md` (or `../pixverse-podcast-clip/SKILL.md`,
   `../pixverse-street-interview/SKILL.md` for those formats), B-roll stills with
   `../pixverse-create-image/SKILL.md`, montages through Seedance reference video,
   product identity from supplied photos. Use `../../skills-shared/prompt-kits.md` wording.
   Real people, logos and screens come from supplied assets, never lookalikes.
3. Align the takes and render with `../../skills-shared/word-timing.md` and
   `../../skills-shared/anchored-composition.md`; rebuild boards, stickers and titles as
   `../../skills-shared/graphics-components.md` states bound to the new anchors. Reuse the
   reference's music only when the user owns it; otherwise generate or supply a bed.

Keep the reference file, notes and evidence in the project. They are not delivered; the new
video and its editable script, timeline, plan and graphics are.

## Finish And Vary

Watch the finished file against the reference's relationships: does the reveal land on its
word, does the board keep its state, is the caption readable, does the music yield to speech?
Fix the script or plan and re-render before regenerating any take. For a family of versions
(new host, new product, new language, new aspect) continue with
`../pixverse-video-variants/SKILL.md`, which reuses every accepted asset.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-video-restyle
description: "Video restyle: change the look, world, wardrobe, subject or medium of an existing video while keeping its motion, performance, timing and cuts, through Seedance edit and reference routes plus local grading."
---

# Video Restyle

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

Keep what moves; change how it looks. A restyle preserves the source's duration, motion,
performance, camera and cuts while replacing its style, setting, wardrobe, medium or the
subject's appearance. It is not a remake and not a new clip that merely resembles the source.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Decide What Survives

Inspect the source once with `"${PVX}" media probe` and a contact sheet (`media tile
--every 1`). Record duration, aspect, cut times, visible text, speech and which elements are
protected. Write a change sheet: one target change per output, protected elements, and the
acceptance checks. Distinguish:

- **Local grade, crop, retime, overlay**: `../pixverse-video-editing/SKILL.md`, no generation.
- **Look/world/wardrobe/medium change with the same motion**: the generative routes below.
- **Different person doing the same performance**: identity replacement, also below, with a
  supplied or accepted reference person.
- **Motion transferred onto a new character**: `create motion-control` (character image plus
  motion video); it does not preserve the source scene, voice or edit.

Explain which route is possible before spending; a candidate route is not proof of exact
preservation. Read `./references/routes.md` for the current capability walls.

## Generate

Compose the queue through the gateway. The primary route is Seedance 2.5
`create reference --task-type edit` with the source in `--videos`, `--duration auto`,
`--aspect-ratio auto`, `--quality 1080p`, plus identity/style references in `--images`.
Write the prompt as a preservation contract: what the model must keep (every movement,
gesture, cut, framing, timing, untargeted text), then the one change, then the render style
(`../../skills-shared/prompt-kits.md` for wording). Each output binds the original source,
never a previous variant. For long sources split by scene with `media cut` and rejoin.

The CLI has no Seedance audio toggle. Plan to restore the original audio at export when
speech or music must survive; changed lip motion can invalidate reuse of original speech.

## Compare And Deliver

Compare source and result on the same contact-sheet grid (`media tile --every 0.5` over the
difficult motions, cut boundaries and text moments). Check full duration, motion fidelity,
identity or style consistency across cuts, and that untargeted text survived. Restore the
soundtrack with FFmpeg when authorized. Report any preservation failure honestly instead of
calling a similar-looking clip a restyle. Repair only the rejected version; accepted outputs
stay. For a batch of looks over one source, use `../pixverse-video-variants/SKILL.md`.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-video-script
description: "Video script and YouTube copy: write the full spoken body with a clear hook, scene beats, visual cues and a usable ending. Script-only work needs no media generation."
---

# Video Scripts

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Deliver words someone can actually speak. An outline is useful preparation, not the finished script unless requested.

For a complete YouTube/hosted script, read `./references/genre-routing.md` first and then
exactly one selected genre guide in full. Deliver three hooks, a complete timestamped
spoken body with production cues, and a retention map. Route by viewer promise/evidence,
not topic nouns; preserve each genre's short-form rules. Ads and fictional scenes retain
their own specialized workflow instead of borrowing an unrelated YouTube beat map.

## Shape The Argument Or Story

Establish audience, desired duration, central question and available evidence. Infer the tone from the brief; a practical tutorial, personal commentary and historical documentary need different language.

Choose the structure that serves the subject:

- Explanation: question, concrete example, mechanism, implication.
- Review: use context, observed evidence, trade-offs, who it suits.
- Commentary: position, reasons, strongest objection, conclusion.
- Documentary/history: sourced event, cause, turning point, consequence; distinguish reconstruction from fact.
- Creator ad: need, visible product proof, practical demonstration, appropriate invitation.
- Narrative: a character wants something, faces an obstacle and makes a consequential choice.

Use sources for factual claims. Avoid fabricated quotations, testimonials, personal experience and unsupported health or performance claims. Keep provisional facts out of finished sales copy.

## Write The Full Body

Write the opening, all spoken sections and the ending. Use contractions or conversational phrasing appropriate to the requested language. Give transitions enough context to be spoken naturally. Maintain a pronunciation list for names, brands and mixed-language terms.

Separate spoken text from visual intentions and source notes. A visual intention states what the viewer must understand; it need not dictate a lens for every sentence. Estimate duration from language and delivery, then label the estimate; actual recording or TTS will set timing.

## Hand Off

For script-only work, deliver the complete editable text and unresolved factual questions. For production, the parent workflow uses this text as its canonical spoken source. Do not add image/video tasks or ask for account login to finish a script.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 8

View saved version →

---
name: pixverse-video-translate
description: "Video localization: deliver the same video in another language with a rewritten script, native or narrated speech, re-timed captions and graphics, and typography that fits the target script."
---

# Video Localization

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

The same piece in another language, not a subtitle pass. The script is rewritten for the
target audience, the performance is spoken in that language, and captions, boards, labels
and sounds re-flow to the new words because they are bound to script anchors.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Translate The Meaning

Rewrite with `../../skills-shared/semantic-script.md`, keeping the same segments and anchor
names so the plan needs no change. Translate punchlines and claims for effect, not word by
word; keep numbers, names and product terms exact with `<shown|said>` readings. Cue breaks
follow the target language's reading rhythm: Chinese and Japanese phrases by meaning without
spaces, Spanish and German with longer phrases and fewer, larger cues. Measure the new
segments; translations change length, so re-choose take durations.

## Produce The Speech

- **Native speech** (default when a person is on camera): regenerate each take through
  `../pixverse-talking-head/SKILL.md` with the same host image and a voice reference in the
  target language (`create voice` with a matching preset and `--language`).
- **Narration over existing picture**: keep the accepted takes muted, generate narration with
  `../pixverse-voiceover/SKILL.md`, and align it as an audio take so anchors come from the
  narration. In the render plan use `base: {"media": "<absolute original video path>",
  "mute": true}` and an `audio` layer pointing to the narration with `during: "program"`.
  Recut or explicitly extend the picture to the measured narration duration before export;
  a longer narration does not automatically extend the original picture.
- **Dubbing an existing recording**: only when the user accepts that lips will not match;
  otherwise choose native regeneration.

Reuse everything that does not speak through `../pixverse-video-variants/SKILL.md`.

## Re-flow And Typeset

Rebuild the timeline with `--language <code>` (`../../skills-shared/word-timing.md`), then
re-render the plan (`../../skills-shared/anchored-composition.md`). Captions pick a CJK face
automatically; check glyph coverage and line length for the target script, and pass
`--size`/`--margin-v` when a longer language needs it. Re-render labels, stickers and board
text with the translated strings (`../../skills-shared/graphics-components.md`). Read
`./references/language-notes.md` for per-language caption and timing rules.

Deliver one file per language with its script, timeline and plan. Sample the same anchors in
every language to confirm reveals still land on the right words.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-video-variants
description: "Video variants: produce many versions of one generated video project by swapping the host, product, wording, language, aspect or hook while reusing every accepted asset and regenerating only what changes."
---

# Video Variants

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.

One editable production, many deliverables. A variant changes one or a few roles (host,
product, wording, language, aspect, hook) and reuses everything else: accepted images,
takes, music, graphics, timeline and plan. Paid work is only the parts that actually change.
Ordered edits of a supplied ad stay in `../pixverse-ad-variants/SKILL.md`.

## Installed Command

Before running helper commands, resolve the absolute plugin root from this `./SKILL.md` path: it is two directories above this skill directory and contains `.codex-plugin/plugin.json`. In every new shell session, set:

```bash
PVX="<absolute-plugin-root>/scripts/pvx"
```

Run helper commands through `"${PVX}"`. Do not assume the user's current working directory is the plugin checkout, and do not bypass the installed wrapper with a module invocation.

## Define The Variant Sheet

Start from an existing project with its script, timeline, plan and manifest. For each
variant record: the change (what and where it appears), the assets that stay, the assets to
regenerate, and the acceptance check. Zip ordered lists; never build a Cartesian product the
user did not ask for. Read `./references/reuse-matrix.md` for what each change touches.

## Reuse Before Generating

Carry accepted assets into the new queue without paying again:

```bash
"${PVX}" queue append projects/<slug>-v2/queue.json --project <slug>-v2 --id broll-leaves --reuse <slug>:broll-leaves
"${PVX}" queue append projects/<slug>-v2/queue.json --id host-cat --reuse <slug>:host-cat
"${PVX}" queue append projects/<slug>-v2/queue.json --id take --preflight -- pixverse create reference --model seedance-2.5 ... --images "{{host-cat.path}}" --prompt ...
```

`--reuse <project>:<task-id>` copies the accepted asset's provider path and local file from
that project's manifest; the preflight lists it as reused with no charge and dependents can
reference `{{id.path}}` as usual. Generate only the changed nodes: a new host image and its
takes for a host swap, new takes for new wording, a new voice and takes for a language,
new inserts for a product. Graphics with the swapped identity (an avatar in a sticker, an
icon on a board) are re-rendered locally, not regenerated.

The example assumes `host-cat` is already an accepted source-project asset; otherwise
generate the new host image before referencing it. The identifier after the colon is the
queue task key, not a provider generation ID. The helper selects its latest successful
manifest entry; it cannot infer creative acceptance or rejection. Verify the selected task
ID against project decisions before reuse, and use a distinct task key for rejected retakes.

## Re-flow The Composition

Rebuild the timeline from the new takes (`../../skills-shared/word-timing.md`); anchors,
captions, boards and sounds follow the new words. Re-render the plan
(`../../skills-shared/anchored-composition.md`) with only the changed sources swapped. For
an aspect change, keep the script and takes when the framing allows it and adapt the plan
geometry; regenerate takes only when the composition truly needs a different frame.

Timing re-flow does not relocate graphics around a new face, prop or gesture. Inspect the
new take at the overlay anchors and re-place cards, boards, inserts and stamps before the
final render. Reuse the layout only after checking its occupied regions on the new host.

## Deliver As A Set

Sample one grid per variant on the same anchors and compare. Deliver the files with a
compact sheet: variant, what changed, what was reused, task ids and credits from the ledger.
Repair only the rejected variant. Keep the project editable so the next batch starts from
the same accepted base.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Local media, script, timeline and graphics commands spend no credits.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2

View saved version →

---
name: pixverse-visual-recipes
description: "Visual effects: apply a selected image treatment or short motion recipe to the user's subject through original prompts and suitable PixVerse routes."
---

# Visual Recipes

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Translate the requested effect into a small set of physical choices. A recipe guides art direction; it is not a hidden model capability or an executable slash command.

Read `./references/execution.md` for selected treatment/preset resolution, browsing
versus execution, exact inputs, one submission and ordered result continuation.

## Choose The Treatment

First retain the requested medium, count, aspect, duration and reference. Select only a relevant family:

- Named product image treatment: `./references/preset-images.md`
  (45 recipes, one image and exact composition/text rules).
- Named product motion treatment: `./references/preset-motion.md`
  (18 recipes, required prepared start frame followed by a silent 6-second video by default).

- Image composition, lighting, material effects and everyday scenes: `./references/image-directions.md`.
- Camera movement, object rotation, lighting change and moving materials: `./references/motion-directions.md`.
- Exact product information or comparison: `../pixverse-listing-images/SKILL.md`.
- Content-linked thumbnail or cover: `../pixverse-cover-art/SKILL.md`.

Use the family to decide subject placement, a specific change, reference needs and likely failure. Combine effects only if the combined action remains legible within the requested duration. A free short motion brief stays a direct video task unless subject precision requires a control image. A selected product motion recipe requires its named starting-plate stage; do not drop it under the general simplicity rule.

## Form The Prompt

Describe the source object and what must remain recognizable. Specify the camera or material behavior with a clear start and finish. For motion, distinguish camera travel, object motion and light motion. Add intended sound in plain language. Keep technical model flags outside the prompt.

Example: “The reference ceramic speaker remains stationary on a charcoal surface. Start with a dim rim outlining its left edge. Over six seconds a broad soft light travels from left to right, revealing the woven front grille and the two top controls. The camera holds the same low three-quarter view throughout. Keep the body shape, control count and reference mark intact. Quiet room ambience; no speech or music.”

This example is a direction to adapt, not a stock prompt to append to every request.

## Add A New Direction

Add an original row to the appropriate reference with five useful parts: user effect, composition/action, required evidence, failure check and cheapest repair. Give a new top-level skill only to a distinct deliverable that needs its own decisions. Keep examples self-authored and model-independent. Do not invent provider preset IDs.

Verify the visible effect against the request. A camera orbit is not proof that the object rotated; a plausible back view is not an accurate unseen product design. If complete geometric fidelity matters, require matching product views and keep unsupported unseen details unresolved.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 6

View saved version →

---
name: pixverse-voiceover
description: "Voiceover: produce natural narration or explicitly timed takes, preserving pronunciation, voice continuity and measured timing for editing."
---

# Voiceover

When this workflow generates images or video, apply `../../skills-shared/quality-policy.md`:
Sunburst 2K/high for images, Seedance 2.5 1080p with automatic prompt enhancement for video,
and a stopped upgrade/fallback choice for Free/Basic or model-entitlement rejection.


Start with the words, their pronunciation and the listening context. Fit the delivery to the meaning before optimizing duration.

Read `./references/production.md` for fixed-window takes, a continuous measured story read,
or an existing video plus photo becoming an on-screen narrator. These are distinct modes
with different timing/retry contracts; ordinary TTS and native UGC speech do not inherit
them automatically. Preserve voice locks and accepted takes; report missing revoice or
compositing capability instead of substituting a static portrait.

## Make The Takes

- Preserve the accepted text and speaker language. Choose a voice using the current CLI voice catalog; use real returned IDs. A stored voice ID is not evidence of voice-cloning support.
- Split at natural thought or breath boundaries. Keep continuous passages together when tone matters; one subtitle per TTS request is only useful for a deliberately timed take list.
- Use the voice model's actual supported controls for speed, emotion or style. Send the spoken text separately from nonspoken direction; do not accidentally speak labels such as “pause” unless the engine supports an established markup syntax.
- Submit via the existing paid queue contract. Measure returned duration and listen for word errors, clipped ends and unnatural pauses.
- If a time window is strict, record its bounds and actual voiced interval. Use the
  selected mode's content/duration correction; never time-stretch or cut speech to fit.
  A separately requested speed effect is an explicit edit, not a narration repair.

## Cue Example

Spoken text: “Open the lid, then press the button on the side.”
Direction: reassuring instructional delivery; a natural breath after “lid”; pronounce every action clearly. Keep this direction outside the literal TTS text when the selected model has no markup/control for it.

## Use With Video

Actual audio determines caption and edit timing. Use `../pixverse-captions/SKILL.md` when burned captions are requested. New narration over an on-camera speaker requires verified sync capability or a new native speech shot; do not silently replace the audio beneath old lip motion.

## Deliver

Provide the audio, exact spoken text, measured timing and take identifiers. Only repair failed takes. Avoid introducing a separate narration task where an accepted native video soundtrack already meets the brief.

## Execution

Read `../../skills-shared/production-brief.md` for reference roles, prompt construction and revisions.
For media execution, follow `../../skills-shared/cli-workflow.md`; load the gateway before manual
CLI commands or queue specs. Resolve the plugin root from this skill's installed path:
set `PVX="<absolute-plugin-root>/scripts/pvx"` and invoke `"${PVX}"` with the documented arguments.
Prompt-only work ends with the usable plan and prompts; it needs no account or generation calls.

Referenced files: 2