fal

View saved version →

---
name: character-design
description: >
  Build consistent character designs and character media with fal.ai. Use for
  original characters, reference sheets, expression sheets, outfit variations,
  identity-preserving edits, and character-to-video workflows.
---

# Character Design

Use when the user wants to create, refine, or preserve a character. Execution
runs through fal.ai MCP tools.

The main objective is consistency: keep the anchor stable and vary only the
requested scene, expression, outfit, camera, or action.

## Inputs To Collect

- Character type: realistic, stylized, anime, mascot, fantasy, sci-fi.
- Identity anchor: age range, face shape, hair, eyes, build, posture, marks.
- Style: photo, 3D, illustration, manga, comic, game concept.
- Needed outputs: portrait, full body, turnaround, expressions, outfits, video.
- References: source image, approved design, costume, pose, style board.
- Consistency level: exploratory, pitch-ready, production continuity.

## Anchor System

Create a reusable anchor paragraph:

- Face and head shape.
- Hair and color.
- Eye shape/color.
- Body/build/posture.
- Clothing silhouette.
- Distinctive marks/accessories.
- Style and rendering constraints.

Repeat the anchor in every generation/edit/video prompt. Add only the changing
variable after the anchor.

## Production Routes

- First character concept: premium still image route.
- Reference sheet: one prompt skeleton, multiple poses/expressions.
- Outfit variants: preserve face/body anchor, vary only clothing.
- Scene/action variants: reference/edit route when identity must hold.
- Character video: approved still first, then image-to-video with anchor.

## Quality Bar

- Face, hair, body proportion, and style remain stable.
- Outfit or pose changes do not rewrite identity.
- Video preserves the approved first frame.
- Final answer includes anchor text so future runs can reuse it.

View saved version →

---
name: cinematography
description: >
  Design cinematic image and video prompts for fal.ai. Use for shot language,
  camera movement, lighting, lens choices, color grade, scene blocking,
  film texture, and production-ready visual direction.
---

# Cinematography

Use when the user needs concrete visual direction, not generic "cinematic"
prompting. Execution runs through fal.ai MCP tools.

## Inputs To Collect

- Subject and action.
- Medium: still, video, image-to-video, edit, storyboard frame.
- Genre and mood.
- Framing: close-up, medium, wide, overhead, POV, profile, locked-off.
- Camera motion: push-in, dolly, tracking, handheld, crane, drone.
- Lens feel: wide, normal, telephoto, macro, shallow/deep focus.
- Lighting: natural, practical, studio, noir, high key, low key, backlit.
- Output: aspect ratio, duration, first frame, last frame.

## Prompt Build

1. Subject/action.
2. Shot size and camera position.
3. Camera motion over time if video.
4. Lens/depth behavior.
5. Lighting source and direction.
6. Color palette and grade.
7. Texture: clean digital, film grain, archival, phone, broadcast.
8. Constraints: no identity/product drift, no fake text, no extra logos.

## Model Choices

- Premium still: `openai/gpt-image-2`, `fal-ai/nano-banana-pro`.
- Final video: Seedance 2.0 text/image video.
- Fast video draft: Grok Imagine Video.
- Multi-prompt/control-heavy shot: Kling v3.

## Quality Bar

- Prompt describes an actual camera setup.
- Motion is physically plausible.
- Lighting direction matches the setting.
- Video prompt has a beginning and ending state.
- Any referenced first frame is treated as an anchor, not redesigned.

View saved version →

---
name: commercial
description: >
  Plan and produce commercial image or video assets with fal.ai. Use for
  product photography, ads, e-commerce batches, product reveals, lifestyle
  commercials, background replacement, platform crops, and brand-safe prompts.
---

# Commercial Production

Use when the user wants advertising, product, brand, or e-commerce media.
Execution runs through the fal.ai MCP tools; this skill provides production
judgment, not another command layer.

Load `model-routing` when the model is not specified.

## Inputs To Collect

- Product: name, category, material, color, scale, logo/packaging rules.
- Goal: hero shot, PDP image, ad creative, product reveal, demo, lifestyle.
- Platform: square, vertical, landscape, banner, transparent background, print.
- Brand: premium, playful, clinical, athletic, minimal, natural, technical.
- Source media: packshot, logo, reference scene, prior generated asset.
- Constraints: preserve packaging, avoid new labels, no fake readable copy.

## Production Workflow

1. Choose the route: text-to-image, reference edit, product cleanup, or
   still-to-video reveal.
2. Use reference/edit workflows when product fidelity matters.
3. Build an approved still before video unless the user explicitly wants
   text-to-video exploration.
4. For e-commerce batches, keep one prompt skeleton and vary one axis at a
   time: background, crop, lighting, platform, or color treatment.
5. For text overlays, reserve clean space unless using a text-capable model
   with exact user-provided copy.

## Prompt Order

1. Product invariant: object, material, packaging, scale, logo rules.
2. Commercial role: hero, PDP, launch teaser, social ad, demo.
3. Setting: surface, background, props, environment.
4. Lighting: softbox, strip light, rim light, practicals, caustics.
5. Camera: angle, focal-length feel, macro, depth of field, motion.
6. Composition: safe zones, crop, negative space, platform.
7. Brand tone.
8. Guardrails: preserve logo/package, no extra text, no distorted labels.

## Model Choices

- Text-heavy ads/posters/UI: `openai/gpt-image-2`.
- Premium product stills: `openai/gpt-image-2`, then
  `fal-ai/nano-banana-pro`, then `fal-ai/nano-banana-2`.
- Product/reference edits: `fal-ai/nano-banana-pro/edit`, then
  `openai/gpt-image-2/edit`.
- Product reveal video: `bytedance/seedance-2.0/image-to-video`.
- Fast draft video: Grok Imagine Video endpoints.

## Quality Bar

- Product shape, logo, material, and color are stable.
- Generated text is absent or exact and intentional.
- Composition supports crop and external copy.
- Props support the product instead of competing with it.
- The final answer includes endpoint, request id when available, output URL,
  prompt summary, and any defects.

View saved version →

---
name: fal-gamedev
description: >
  Generate 2D game assets with fal.ai. Use for pixel-art characters, sprite
  sheets, walk cycles, RPG sprites, idle/attack/jump animations, transparent
  assets, parallax backgrounds, isometric maps, and game-ready visual packs.
---

# fal.ai Game Dev Assets

Use when the user wants 2D game art or asset packs. Execution runs through
fal.ai MCP tools.

## Asset Pipeline

1. Character concept: generate or edit a clean base character.
2. Sprite poses: create walk, idle, jump, attack, side/up/down/isometric views.
3. Cleanup: remove background and preserve transparent output when needed.
4. Backgrounds: generate parallax layers or isometric map assets.
5. Consistency pass: ensure palette, line weight, proportions, and camera angle
   match across all assets.

## Inputs To Collect

- Game genre and camera: platformer, RPG, isometric, top-down, side-scroller.
- Style: pixel art, hand-painted, vector, anime, low-poly render.
- Character description and palette.
- Needed animations and directions.
- Tile/grid constraints, transparent background, output size.
- Engine constraints if supplied.

## Prompt Guardrails

- Keep one character silhouette across all sprites.
- Specify camera angle and frame count expectations.
- Avoid text unless UI asset text is explicitly requested.
- Use consistent palette and lighting across the pack.
- For sprite sheets, reserve spacing between poses.

## Quality Bar

- Assets are readable at game scale.
- Directional sprites match perspective.
- Transparent assets do not leave halos when background removed.
- Background layers can parallax without obvious seams.

View saved version →

---
name: fal-media
description: >
  Generate, edit, or process media through fal.ai. Use when the user names Fal
  or fal.ai, requests Fal model discovery, schema, pricing, or job handling, or
  asks to use a named media model available through Fal such as MiniMax H3 Max.
  Do not activate when the user explicitly chooses another provider.
---

# fal.ai Media

Route media generation and processing through the fal.ai MCP tools exposed by
this plugin. This is the provider-level entry point; load a more specialized
Fal skill when the task needs its production guidance.

## Provider And Model Routing

- Preserve an explicitly requested provider. When the user names Fal, keep the
  operation on Fal. When the user names another provider, do not invoke Fal.
- For a named model, search Fal's live catalog and verify its exact canonical
  endpoint. Prefer an exact match over a similarly named model exposed by an
  intermediary. Do not silently substitute a different model or provider.
- When no model is specified, call `recommend_model` for a current candidate,
  then judge the result against the requested artifact and constraints.
- Inspect the selected endpoint schema before constructing its input. Check
  pricing when duration, resolution, batching, training, audio, video, or 3D
  can materially affect cost.

## Execution

- Use `run_model` for bounded image and utility operations.
- Use `submit_job` for video, audio, 3D, training, or other long-running work;
  poll the same request with `check_job`, then fetch it with `get_job_result`.
- Upload local inputs before generation when the endpoint requires a URL.
- Preserve continuity with image-to-video, reference-to-video, or frame
  extraction when an earlier result should anchor the next shot.
- Return the endpoint ID, request ID, output URL, and material limitations.

## Related Skills

- Load `model-routing` for model selection beyond a direct exact match.
- Load `fal-models-catalog` for broader family or modality comparisons.
- Load `fal-prompting` after selecting an endpoint that needs family-specific
  prompt construction.
- Load `fal-workflow` for model chains, batches, or dependent generation steps.

View saved version →

---
name: fal-models-catalog
description: >
  Navigate fal.ai model families by media modality and production role. Use
  when the user asks which model, endpoint, or model family is appropriate for
  image, video, audio, 3D, editing, training, restoration, try-on, or analysis.
---

# fal.ai Models Catalog

This is a curated routing layer for MCP clients. It should guide endpoint selection,
then execution should happen through the fal.ai MCP tools exposed by this
plugin. Do not shell out to genmedia CLI.

Use `model-routing` first for common production defaults. Use this skill when
the question is broader: comparing modalities, finding a category, choosing a
family, or explaining tradeoffs.

## Modality Map

- Text to image: campaign visuals, product stills, character concepts,
  editorial photography, posters, UI mockups, image typography.
- Image to image: edits, inpainting, reference preservation, style transfer,
  product placement, background replacement, upscaling, restoration.
- Text to video: cinematic clips, product reveals, narrative shots, social
  concepts, motion drafts.
- Image to video: animate an approved still, product hero motion, b-roll,
  talking-head source frames, first-frame continuity.
- Reference to video: stronger continuity from characters, products, or style
  references where supported.
- Text/audio to talking head: spokesperson, UGC creator, avatar, lip sync.
- Text to audio: narration, TTS, music, sound effects.
- Audio to text: transcription, subtitles, diarization, audio cleanup.
- Image/text to 3D: objects, characters, game assets, GLB/OBJ/PLY outputs.
- Image to text / vision: OCR, captioning, segmentation, detection, analysis.
- Training: LoRA and fine-tune style workflows, only when a dataset exists.

## Selection Pattern

1. Identify the artifact role: final commercial, draft, utility transform,
   analysis, training, or intermediate step.
2. If the user did not name a specific endpoint, call `recommend_model` before
   execution and use catalog search when the recommended list is too generic.
3. Choose modality and endpoint family.
4. Inspect schema before assuming fields such as `image_url`, `image_urls`,
   `reference_image_url`, `duration`, `aspect_ratio`, `seed`, `quality`,
   `audio_url`, or `enable_rigging`.
5. Check pricing when the job is long, high resolution, batched, video, audio,
   3D, or likely to be repeated.
6. For uncertain categories, use MCP catalog search and docs search, then
   choose from verified endpoints.

## Production Defaults

- Text-heavy stills: `openai/gpt-image-2`.
- Product stills and campaign heroes: `openai/gpt-image-2`,
  `fal-ai/nano-banana-pro`, `fal-ai/nano-banana-2`.
- Product/reference edits: `fal-ai/nano-banana-pro/edit`,
  `openai/gpt-image-2/edit`.
- Final-quality video: Seedance 2.0 text/image/reference video endpoints.
- Fast video drafts: Grok Imagine Video endpoints.
- Talking head: `veed/fabric-1.0`, `veed/fabric-1.0/text`,
  `fal-ai/creatify/aurora`, `fal-ai/sync-lipsync/v2`.
- Background removal: use catalog search for current Bria/background endpoints
  and inspect schema.
- 3D: prefer Meshy v6 class endpoints for rigging/animation when available.

## Avoid

- Choosing a model only because it is popular when the artifact role is clear.
- Using text-to-image when the user supplied a reference that must be preserved.
- Using cheap draft endpoints for final brand/product assets.
- Generating readable legal, medical, financial, or claim text unless the user
  supplies exact wording and the chosen model supports text well.

View saved version →

---
name: fal-prompting
description: >
  Apply model-family-specific prompting for fal.ai endpoints after an endpoint
  has been selected. Use for GPT Image 2, Kling, Happy Horse, Seedance, product
  prompts, exact text, multi-prompt video, and prompt debugging.
---

# fal.ai Prompting

Use this after model selection. The goal is to write prompts that match the
model family, not generic "cinematic masterpiece" text.

## Universal Rules

- Visual facts beat prestige adjectives.
- Preserve one controlled variable per iteration.
- Inspect schema before assuming negative prompts, seeds, multi-image inputs,
  multi-prompt arrays, duration, audio, or camera controls exist.
- Keep exact text short and quoted when the model supports text.
- For product and character work, name the invariant explicitly in every prompt.
- For video, describe action over time, camera motion, framing, and ending.

## GPT Image 2

Use for exact text, packaging, posters, UI, premium stills, and complex image
composition.

Prompt structure:

1. Output type and composition.
2. Exact subject/product/character.
3. Style and medium.
4. Lighting, material, camera, color.
5. Text requirements, only if needed.
6. Guardrails: no extra logos, no unreadable text, preserve reference details.

GPT Image 2 benefits from structured, explicit prompts. It can handle longer
instructions than fast video models.

## Kling

Use for controlled video prompts when schema supports multi-prompt or element
controls.

Prompt structure:

1. Subject identity and first frame.
2. Shot type and camera motion.
3. Action sequence in temporal order.
4. Motion constraints.
5. Ending state.

Keep control fields schema-driven. Do not invent element fields without schema.

## Happy Horse

Use short natural-language prompts. Long production paragraphs usually hurt.

Good pattern:

`Handheld phone video, close shot of a runner tying neon shoes on wet pavement, soft morning light, quick push-in.`

Avoid stacking style tags or long guardrails. Put important action first.

## Seedance

Use for high-quality cinematic image-to-video and text-to-video. Build prompts
like a short shot direction:

- Subject/action.
- Camera motion.
- Environment.
- Lighting.
- Duration feel.
- What should not change.

For image-to-video, the uploaded frame is the anchor. The prompt should animate
the frame, not redesign it.

## Prompt Failure Fixes

- Product changed: switch to reference/edit workflow and tighten invariants.
- Text is wrong: use GPT Image 2, shorten text, or reserve safe space for
  external typography.
- Video drifts: reduce prompt scope, use image-to-video from an approved frame,
  shorten duration, or choose a model with stronger controls.
- Character identity drifts: use the character anchor wording and reference
  images in every step.
- Output is generic: replace style adjectives with concrete camera, light,
  materials, setting, and action.

View saved version →

---
name: fal-recipes
description: >
  Use-case-driven fal.ai production recipes. Trigger for commercials, product
  photography, cinematic shots, character design, storyboards, talking heads,
  lip sync, virtual try-on, restoration, video with audio, realism, and other
  multi-step media outcomes.
---

# fal.ai Recipes

Recipes are production playbooks. They decide what to collect, what model
families to combine, and what quality bar to enforce. Execution stays in the
fal.ai MCP tools.

## Recipe Index

- Commercial/product: load `commercial`.
- Marketing campaign: load `marketing`.
- UGC/talking head: load `ugc`.
- Cinematic still/video: load `cinematography`.
- Character consistency: load `character-design`.
- Multi-shot story: load `storytelling`.
- Fan-cam sports broadcast: load `fan-cam`.
- 2D game assets: load `fal-gamedev`.
- 3D character selector experience: load `fal-regenerate-3d`.
- Endpoint choice: load `model-routing` and `fal-models-catalog`.
- Prompt family tuning: load `fal-prompting`.

## Standard Recipe Shape

1. Identify the final artifact and platform.
2. Collect only inputs that affect model choice or continuity.
3. Choose endpoint families with `model-routing`; when no endpoint was named,
   start with `recommend_model` and then apply the recipe quality bar.
4. Upload references before generation when fidelity matters.
5. Inspect schemas and pricing for the chosen endpoints.
6. Generate in stages: anchor stills first, then edits/video/audio/3D.
7. Use async jobs for video, audio, 3D, batches, and long-running utilities.
8. Review output against the skill quality bar before returning URLs/results.

## Common Multi-Step Patterns

Product ad:

- Reference product image.
- Product-faithful hero still.
- Optional cleanup/background removal.
- Image-to-video reveal from approved frame.
- Final manifest with endpoint, request id, prompt summary, and defects.

UGC ad:

- Script and hook.
- Creator portrait/avatar route.
- Product b-roll route.
- Talking-head or lip-sync generation.
- Optional cutdown variants by hook angle.

Story:

- Beat sheet.
- Shot list with continuity anchors.
- Keyframes or first frames.
- Image-to-video per shot.
- Optional narration/audio.

Character:

- Identity anchor.
- Reference sheet.
- Outfit/expression/action variants.
- Optional image-to-video with identity guardrails.

## Quality Bar

Do not return a recipe as complete until the result matches the requested
artifact role. If the model output misses the role, adjust the route before
adding more variants.

View saved version →

---
name: fal-regenerate-3d
description: >
  Build a polished 3D character-selector experience with fal.ai assets. Use
  for stylized characters, GLB generation, companion objects, PBR floor
  textures, video backgrounds, palette systems, and Three.js presentation.
---

# fal.ai 3D Experience Pipeline

Use when the user wants an end-to-end 3D showcase, roster, selector, or game
character experience. Execution of media assets runs through fal.ai MCP tools;
the final app is built in local code.

## Pipeline

1. Character stills: generate full-body front-facing designs with stable
   silhouettes and palette.
2. Character 3D: convert approved stills to GLB/OBJ/PLY with rigging or
   animation fields when supported.
3. Companion/object: generate matching prop or creature stills, then convert
   to 3D if needed.
4. Materials: generate/extract PBR floor or environment textures.
5. Backgrounds: generate loopable video or still environment per character.
6. App shell: Three.js scene, character selector, palette swaps, responsive UI.
7. Optimization: compress GLBs, reduce textures, lazy-load assets, test mobile.

## Inputs To Collect

- Experience theme and number of characters.
- Character archetypes, palettes, and environment biomes.
- Output format: static HTML, React app, game asset pack, GLB set.
- Rigging/animation needs.
- Budget/quality target and maximum asset size.

## Quality Bar

- GLBs load and render nonblank.
- Materials are not overlarge for the web target.
- Character identity matches source still.
- Selector UI is responsive and usable.
- The final answer includes all generated asset URLs, local files when present,
  and integration notes.

View saved version →

---
name: fal-workflow
description: >
  Design multi-step fal.ai media workflows for MCP execution. Use when the user
  wants model chaining, batch generation, fan-out/fan-in, frame bridging,
  narrated video, dataset creation, reusable pipeline logic, or workflow JSON.
---

# fal.ai Workflows

Use this skill when a single model call is not enough. A workflow is a planned
sequence of model calls with dependencies, retries, quality gates, and a final
manifest.

## Workflow Modes

- Cloud workflow JSON: use when the user wants a reusable graph asset.
- MCP-orchestrated workflow: use when the client should run a sequence now through
  the fal.ai MCP tools.

Do not call genmedia CLI from this plugin. If a source skill mentions CLI
orchestration, translate that intent into MCP tool usage.

## Planning Pattern

1. Define final deliverable and acceptance criteria.
2. Split into nodes: input assets, generation, utility transforms, edits,
   video/audio/3D steps, and final packaging.
3. Choose model per node with `model-routing`.
4. Capture dependencies: which output URL feeds which next input.
5. Inspect schema for each endpoint before building payloads.
6. Run independent expensive nodes async.
7. Poll async nodes without restarting completed work.
8. Gate each step: preserve identity/product, check duration, check format.
9. Return a manifest: node, endpoint, request id, inputs, output URLs, defects.

## Useful Patterns

- Anchor first: generate or upload a strong still, then drive edits/video from
  that approved asset.
- Fan-out: create several variants from the same anchor, then select one.
- Fan-in: combine product, logo, character, or style references into an edit.
- Frame bridge: generate first and last frame, then use image-to-video or
  supported controls to bridge motion.
- Audio-first: generate/upload voice first, then match visuals to duration.
- Utility cleanup: background removal, compression, upscaling, format
  conversion, segmentation, or OCR as deterministic nodes.

## Failure Handling

- 422/schema error: re-inspect schema and remove guessed fields.
- Product/identity drift: switch to reference/edit workflow.
- Video still running: keep polling the same request id.
- Bad prompt alignment: reduce prompt scope and change one variable at a time.
- Cost spike: reduce variants, duration, resolution, or use draft endpoints.

View saved version →

---
name: fan-cam
description: >
  Create personalized live sports broadcast fan-cam videos with fal.ai. Use
  for realistic spectator cutaways, stadium crowd reactions, broadcast-style
  screenshots, scoreboard overlays, TV bugs, and identity-preserving fan
  reaction videos from a user photo.
---

# Fan Cam

Use when the user wants a personalized spectator video that feels like a live
sports broadcast cutaway. Execution runs through fal.ai MCP tools.

## Required Inputs

- User photo or approved portrait reference.
- Sport, team/event context, venue, and crowd mood.
- Reaction: shocked, celebrating, nervous, chanting, laughing, emotional.
- Broadcast format: close-up cutaway, crowd pan, scoreboard moment, replay.
- Duration and crop, usually 9:16 or 16:9.

## Production Pattern

1. Upload the portrait reference if local.
2. Generate a broadcast-still frame that preserves identity.
3. Use image-to-video from that approved frame.
4. Add sports broadcast details in prompt only when they are generic or
   user-provided. Do not invent real network marks or false event claims.
5. Review identity, crowd plausibility, scoreboard/text artifacts, and motion.

## Prompt Guardrails

- Keep the person recognizable from the reference.
- Use generic broadcast styling unless the user supplies exact rights-safe
  graphics.
- Avoid fake readable scoreboards unless the user supplies exact text.
- Keep motion natural: cheering, gasping, standing, waving, phone held up.

## Model Choices

- Identity-preserving frame: `openai/gpt-image-2/edit` or another strong edit
  endpoint after schema inspection.
- Broadcast video: Kling v3 image-to-video or Seedance image-to-video,
  depending on available controls and quality target.

## Quality Bar

- Person remains recognizable.
- Broadcast frame feels live, not posed studio content.
- Crowd and venue support the sport.
- Any text/logos are absent, generic, or explicitly provided.

View saved version →

---
name: marketing
description: >
  Plan campaign-level marketing asset production with fal.ai. Use for launch
  kits, campaign matrices, paid social variants, landing-page visuals, email
  and banner imagery, hook/proof/conversion assets, and creator ad packages.
---

# Marketing Production

Use when the user wants a campaign system rather than one isolated asset.
Execution runs through fal.ai MCP tools.

Load `commercial` for product-specific assets and `ugc` for creator-style ads.

## Inputs To Collect

- Objective: launch, acquisition, retargeting, education, event, activation.
- Audience: persona, market, use case, awareness level, objections.
- Offer: product, feature, bundle, trial, waitlist, event, promotion.
- Channels: paid social, organic, landing page, email, display, app.
- Required assets: stills, videos, thumbnails, hero image, carousel, banner.
- Brand rules: colors, logo, typography, tone, taboo visuals, competitors.
- Claims: exact approved copy, proof points, disclaimers, compliance limits.

## Campaign Matrix

Create a compact matrix before generation:

- Hook asset: first-frame attention.
- Proof asset: product, process, visible result, or supplied evidence.
- Context asset: audience use case.
- Conversion asset: clean end frame or layout-safe visual.
- Retention/reminder asset: variation for later touchpoints.

Each asset needs role, channel, crop, format, model route, prompt summary, and
guardrails.

## Prompt Order

1. Asset role and channel.
2. Product or subject invariant.
3. Audience context and use case.
4. Visual system: camera, lighting, composition, color, motion.
5. Copy handling: no generated text, safe space, or exact provided wording.
6. Variation axis.
7. Guardrails: no fake claims, extra logos, distorted UI, or fake proof.

## Model Choices

- Exact copy/key art: `openai/gpt-image-2`.
- Premium stills: `openai/gpt-image-2`, `fal-ai/nano-banana-pro`,
  `fal-ai/nano-banana-2`.
- Edits from product/logo/UI: `fal-ai/nano-banana-pro/edit`,
  `openai/gpt-image-2/edit`.
- Product/social video: Seedance 2.0 image-to-video or text-to-video.
- Creator ad: `veed/fabric-1.0`, `veed/fabric-1.0/text`, or
  `fal-ai/creatify/aurora`.

## Quality Bar

- Every asset maps to a campaign role and channel.
- Claims are supplied, observable, or removed.
- Variants differ by one clear axis.
- Safe zones work for platform UI and external overlays.
- Final answer includes a campaign manifest and defects.

View saved version →

---
name: model-routing
description: >
  Choose production-ready fal.ai endpoint IDs for MCP media workflows. Use
  with commercial, marketing, ugc, character-design, cinematography,
  storytelling, fal-recipes, and fal-workflow when the user has not named a
  specific model or asks which model to use.
---

# fal.ai Model Routing

Use this skill for curated endpoint choice. It is not a replacement for the
MCP tools. After choosing an endpoint, execute through the fal.ai MCP tools in
this plugin. Do not call the genmedia CLI from this MCP skill bundle.

## Routing Rules

1. If the user did not name a specific endpoint, call `recommend_model` first
   for a current catalog-ranked recommendation.
2. Use the curated endpoints in this skill to evaluate the recommendation and
   pick the artifact-appropriate route. If the recommendation is generic but
   the task is specialized, use `search_models` to verify a better endpoint.
3. Use live MCP discovery when the role is not covered, an endpoint is missing
   or deprecated, or the user asks to compare options.
4. Inspect the selected endpoint schema before using custom fields.
5. Check pricing when cost, batch size, video duration, 3D, or high-resolution
   output matters.
6. Do not invent endpoint IDs or schema fields.

## Image Generation

Generic/simple image generation:

- `fal-ai/flux-2/klein/9b`: default when the user asks for a basic image and
  does not name a model or require premium text rendering, brand fidelity, or
  final commercial quality.

Text-heavy image work:

- `openai/gpt-image-2`: best default for readable text, posters, packaging,
  diagrams, UI mockups, book covers, exact-copy layouts, and high-end stills.
  Use high quality when the result is final.
- `fal-ai/nano-banana-pro`: second choice when text is important but cost or
  availability matters.

Premium still images:

- `openai/gpt-image-2`: commercial realism, editorial photography, product
  scenes, character anchors, and high-quality concept art.
- `fal-ai/nano-banana-pro`: strong styled output and product-friendly images.
- `fal-ai/nano-banana-2`: cheaper strong alternative.

Fast draft images:

- `fal-ai/flux-2/klein/9b`: quick concepts, mood exploration, rough layouts.
  Do not use as final commercial delivery unless the user asks for speed/cost.

## Image Editing

Use for product edits, background replacement, inpainting, relighting, cleanup,
object changes, outfit changes, and reference-driven composition.

1. `fal-ai/flux-2/klein/9b/edit`: default for simple/generic image edits when
   the user does not name a model.
2. `fal-ai/nano-banana-pro/edit`
3. `openai/gpt-image-2/edit`
4. `fal-ai/bytedance/seedream/v5/lite/edit`

For product fidelity, prefer reference/edit workflows over text-only generation.

## Video

Highest quality video:

- Text to video: `bytedance/seedance-2.0/text-to-video`
- Image to video: `bytedance/seedance-2.0/image-to-video`
- Reference to video: `bytedance/seedance-2.0/reference-to-video`

Fast/lower-cost video:

- Text to video: `xai/grok-imagine-video/text-to-video`
- Image to video: `xai/grok-imagine-video/image-to-video`
- Video edit: `xai/grok-imagine-video/edit-video`

Multi-shot or control-heavy video:

1. `bytedance/seedance-2.0/text-to-video`
2. `bytedance/seedance-2.0/image-to-video`
3. `bytedance/seedance-2.0/reference-to-video`
4. `fal-ai/kling-video/v3/pro/text-to-video`
5. `fal-ai/kling-video/v3/pro/image-to-video`
6. `alibaba/happy-horse/text-to-video`
7. `alibaba/happy-horse/image-to-video`

Use Kling when multi-prompt, element controls, or stronger prompt structure are
needed. Use Happy Horse for brief, direct camera-language prompts.

## UGC And Talking Head

- Portrait plus audio: `veed/fabric-1.0`
- Portrait plus text: `veed/fabric-1.0/text`
- Avatar with visual direction: `fal-ai/creatify/aurora`
- Existing footage with new speech: `fal-ai/sync-lipsync/v2`
- Product b-roll: `bytedance/seedance-2.0/image-to-video`
- Fast b-roll draft: `xai/grok-imagine-video/image-to-video`

## Audio

Use live catalog discovery for current TTS, music, sound effect, transcription,
and cleanup endpoints. Audio model availability changes quickly, so inspect
schemas and pricing before execution.

## 3D

- Premium image-to-3D or character assets: use Meshy v6 class endpoints when
  available and inspect rigging/animation fields.
- Hunyuan/Tripo-style endpoints can be alternatives for faster drafts or
  different geometry outputs.
- Always run 3D jobs async and return model URLs plus any texture/material URLs.

## Quality Guardrail

Model choice is only valid if it matches the artifact role:

- Text-heavy deliverables need text-capable image models or external layout.
- Product deliverables need reference/edit workflows when packaging must hold.
- Talking-head deliverables need audio/script length checks.
- Story deliverables need continuity anchors across shots.
- 3D deliverables need output format, scale, material, and rigging checks.

View saved version →

---
name: storytelling
description: >
  Build multi-shot narrative image, video, and audio workflows with fal.ai.
  Use for storyboards, shot lists, multi-prompt video, first-frame/last-frame
  pipelines, social stories, brand films, and sequence continuity.
---

# Storytelling

Use when the user wants a sequence, not a single asset. Execution runs through
fal.ai MCP tools.

Load `cinematography` for shot language and `character-design` when identity
continuity matters.

## Inputs To Collect

- Format: ad, short film, music video, documentary, tutorial, social story.
- Duration and aspect ratio.
- Number of shots or range.
- Main subject, character, product, or location.
- Continuity anchors: character, product, wardrobe, environment, color.
- Source media: first frame, reference image, product shot, audio track.
- Audio needs: narration, music, sound design, transcript, or no audio.

## Story Build

1. Beat sheet: what changes emotionally or informationally.
2. Shot list: one visual purpose per shot.
3. Continuity anchors: what must remain stable.
4. Model route: text-to-video, image-to-video, reference-to-video, or
   talking-head/audio path.
5. Keyframes: create or upload approved first frames when continuity matters.
6. Async execution per shot.
7. Review sequence order, duration, continuity, and artifact defects.

## Shot Prompt Pattern

For each shot:

- Shot number and purpose.
- Subject and action.
- Framing and camera movement.
- Lighting and environment.
- Continuity anchors.
- Ending state.
- What must not change.

## Model Choices

- Final high-quality shots: Seedance 2.0.
- Control-heavy video: Kling v3 when schema supports the needed controls.
- Brief simple motion: Happy Horse or fast video endpoints.
- Product/character continuity: use approved stills or reference-to-video.
- Voice/spoken story: use UGC/talking-head routes and align visuals to audio.

## Quality Bar

- Every shot has a narrative job.
- Continuity anchors are repeated and visible.
- The sequence can be edited together without random style drift.
- The final answer includes shot manifest, output URLs, endpoint IDs, and known
  issues.

View saved version →

---
name: ugc
description: >
  Plan and produce UGC-style creator ads and social videos with fal.ai. Use for
  direct-to-camera creator scripts, talking-head ads, product demos,
  testimonials, founder clips, unboxing, reactions, before-after, faceless
  voiceover, and short vertical social videos.
---

# UGC Production

Use when the user wants creator-style content instead of polished studio ads.
Execution runs through fal.ai MCP tools.

## Inputs To Collect

- Product or offer.
- Format: direct-to-camera, demo, reaction, unboxing, founder, faceless b-roll.
- Speaker: supplied portrait/video, generated avatar, no face, or voiceover.
- Script source: exact script, bullets, offer copy, or ask to draft.
- Platform: TikTok, Reels, Shorts, paid social, landing page, prototype.
- Runtime/crop: usually 9:16; 6-15 seconds for hooks, 15-45 for ads.
- Source media: portrait, product image/video, logo, b-roll, audio.
- Claims: proof supplied, disclaimers, banned phrases.

## Script Shape

1. Hook: concrete tension, result, objection, or curiosity.
2. Context: why the speaker cares.
3. Product moment: visible use, demo, or comparison.
4. Proof: supplied metric, visible result, or sensory detail.
5. Turn: objection answered or before-after.
6. Close: soft CTA or final product frame.

Avoid fake testimonials and unsupported medical, financial, or performance
claims.

## Production Routes

- Portrait plus audio: `veed/fabric-1.0`.
- Portrait plus text: `veed/fabric-1.0/text`.
- Avatar with visual direction: `fal-ai/creatify/aurora`.
- Existing footage with new speech: `fal-ai/sync-lipsync/v2`.
- Product b-roll: `bytedance/seedance-2.0/image-to-video`.
- Hook frames/thumbnails: `openai/gpt-image-2` or `fal-ai/nano-banana-pro`.

## Quality Bar

- First 1-2 seconds have a clear hook.
- Spoken script fits runtime.
- Mouth motion is synced for speaking faces.
- Product/logo/packaging stay stable.
- Captions/text are not hallucinated inside the video.
- Claims are user-supplied or phrased as visible observations.