skills-shared/production-brief.md

# Turn A Brief Into Usable Media

Use the smallest plan that protects the requested result. Keep creative reasoning in the host
agent and execution in the existing PixVerse wrapper. These are craft rules, not another router.

## Before A Prompt

Separate three things: supplied facts, visual direction, and deliverables. Preserve the user's
language, roughness, platform, chosen model, count and format. Infer routine art direction;
ask only for a missing input that changes what can truthfully be made.

For a product, identify shape, material, label, working parts, scale and permitted claims.
For a person, retain face, hair, wardrobe and distinguishing features. Name each reference's
job: identity, product, place, motion, voice or style. A style reference is not evidence of a
product's features. Reuse accepted asset IDs/paths in project memory.

Distinguish an opening-frame lock from general reference conditioning. A required exact start
uses the current image-to-video route; adding a general reference alone does not establish
that frame as the opening. Use the gateway's actual command contract for each binding.

One simple shot stays one requested-medium task. A recurring subject, multi-scene demonstration
or exact packaging action needs a shot list and appropriate control frames. Follow the selected
specialized workflow's required board layout, shot count, refinement and audio structure; do not
remove those stages using this general simplicity rule. A recipe-defined storyboard can be the
sequence reference for one multi-cut clip, with separate identity/product anchors. Outside such
a recipe, use the appropriate per-shot control frame rather than an unrelated contact sheet.
Do not add routine image polishing unless the selected workflow requires that refinement or
an observed defect justifies it.

## Prompt Pack

Write a concrete, self-contained prompt for each actual generation task:

- **Visible subject:** physical details plus the role of supplied references.
- **Action:** what changes, from which state to which state; one coherent beat per shot.
- **Framing and movement:** viewpoint, distance, subject position, camera path and endpoint.
- **Look:** specific light, material, palette and medium; omit empty quality adjectives.
- **Sound:** exact spoken words when relevant, language, delivery, ambience, sound effects and music intent.
- **Keep:** the few details the model must preserve and errors especially likely in this shot.

Do not paste a planning essay, validation rubric, file path placeholder or backend instruction
into the visual prompt. Keep render parameters, asset bindings and post-production text separate.
Use the requested look instead of adding a default luxury or cinematic treatment.

For image layouts, reserve space for exact copy. Use deterministic text/graphic layers when
letter accuracy, data, UI or logo geometry matters. The final export must include those layers;
a blank plate and a suggestion to add text later are unfinished work.

## Sound Is A Deliverable Requirement

An unsupported `audio` parameter means the CLI cannot control that switch; it does not prove a
model lacks native sound. Seedance 2.5 uses provider defaults through the current Create route:
omit `--audio` and `--no-audio` in manual queue commands, keep requested sound in the prompt,
and inspect the returned result. Do not add TTS merely because the switch is unavailable.
If a silent file is required, remove any returned audio track during local export and verify it.

Use natural voice segments for narration. Measure actual speech before placing captions and
visual cuts. A timed script is an estimate until audio exists. Putting new TTS over moving lips
does not solve lip sync. Voice-reference input does not promise verbatim copying or voice training.

## Produce Or Plan

For execution, use `./cli-workflow.md` and the gateway's current capability checks, quote and
confirmation policy. Count each image, video and audio task; no automatic variant pack. Reworks
are new paid tasks, and a download failure is repaired by downloading the existing result.

For an explicit Canvas target, use `../skills-internal/pixverse-agent-canvas/SKILL.md`; do not
substitute local helpers for its graph, regional, approval or cloud-preview contracts.

If only prompts or a plan are requested, stop at the concrete prompt pack, reference bindings,
proposed tasks and expected deliverables. No login, quote, generation or dependency installation
is needed to write it. Mark any unavailable input honestly instead of inventing a real asset ID.

## Finish And Revise

Show usable media as it becomes available, then inspect the full requested outcome. Check technical
properties with the existing QA helpers; also inspect product/identity fidelity, action, speech
and exact text. An audio stream alone does not prove audible or correct speech. Mark uninspected
properties as unchecked; do not claim output quality from a prompt review.

Keep accepted assets. A caption color edit changes rendering; an on-camera line change may require
a new video/sync task. Update only dependent shots and the affected timeline. Finish with the actual
files or explicitly selected Canvas result and the existing project handoff.