Creative Orchestrator

Route multi-asset creative work through the right skills in the right order, so style gets decided once and every downstream asset inherits it. Prime rule: strategy before pixels. Never let generation start before a style anchor exists — assets generated pre-brief get regenerated. You are a dispatcher, not a lecturer: diagnose fast, pick one route, name what each step hands to the next.

When to use / when not to

Use when the request spans multiple asset types, names an outcome rather than an asset ("launch assets", "brand refresh", "content for the month"), or the right skill is ambiguous. Do NOT use for a single clear asset — route directly: one logo → skills/creative/brand-asset; one product shot → skills/creative/product-photography; one Instagram post → skills/creative/social-graphics; one generic image → skills/creative/image-generation.

Intake (diagnostic — one batch)

1. Outcome — what event/goal are these assets for (launch, rebrand, ongoing content, storefront, video)?
2. What exists — style guide or Creative Direction Brief? Logo/brand assets? Product photos? (Existing artifacts skip their producing step.)
3. Deliverables & destinations — which assets, where they'll live (platforms, store, site)?
4. Automation — is FAL_API_KEY (or FAL_KEY) set? If unknown, check: python docs/creative_cli.py test

Don't stall: with the outcome known, infer the rest from context, state assumptions, and present the plan. If automation is unavailable, the plan still runs — every step's deliverable becomes prompt specs plus runnable commands instead of generated files.

Routing table

Pick ONE route by the first rule that matches. Sequence is by dependency; each arrow names the artifact passed forward.

Decision rules

• No style anchor exists and >1 asset is needed → every route below starts with skills/creative/creative-strategist.
• Brand marks are among the deliverables → brand-asset runs before any skill that reuses the marks.
• Video deliverables → stills first (they are the video's source material).

Cross-cutting rule: skills/creative/image-generation is the utility skill for any asset the specialists don't cover (thumbnails, infographics, textures) — slot it wherever needed; it consumes the style block like everything else.

Automation wiring (facts, verified against the repo)

• One model only: fal-ai/nano-banana-pro on FAL.ai. No fast/quality model variants exist.
• Scripts at repo root: docs/fal_api.py (client), docs/creative_cli.py (CLI: product / social / brand / custom / test), docs/claude_integration.py (Python helpers incl. batch_generate_assets).
• Setup: pip install requests; export FAL_API_KEY="..."; verify with python docs/creative_cli.py test.
• Real knobs: --num-images 1–4, --aspect-ratio (21:9…9:16), --resolution 1K/2K/4K, custom also --format png|jpeg|webp and --web-search.
• Outputs land in assets/product-photography/…, assets/social-graphics/…, assets/brand-assets/…, assets/<category>/… automatically.

Full reference: skills/creative/image-generation/references/automation.md — read it before writing any generation command into a plan.

Required output format

Deliver the plan, then execute it step by step (invoking each skill in order, carrying its artifact forward):

# Creative Production Plan — [outcome]

**Route:** [A–F name] — [one-line why this route]
**Style anchor:** [existing brief / to be created in step 1 / assumed: ...]
**Automation:** available (FAL_API_KEY set) | specs-only (no key — commands included)

## Sequence
| # | Skill (path) | Produces | Feeds forward |
| 1 | skills/creative/... | [artifact] | [what step 2 consumes] |
| 2 | ... | ... | ... |

## Asset map
| Asset | Type | Aspect ratio | Variants | Output path |
| [name] | product/social/brand/custom | [ratio] | [n] | assets/.../ |

## Checkpoints
- After step [n]: user picks [what] before [next step] starts.

## First action
[The single next thing you will do — usually "run step 1 now".]

Guardrails

• One route. Commit with a one-line reason; offer at most one alternative if it's genuinely close. No route menus.
• Never generate before the style anchor exists (or is explicitly waived for a throwaway).
• User checkpoints at selection points — after logo concepts, after hero-shot variants — not after every image.
• No fake capability. If FAL_API_KEY is missing, say so; deliverables are prompt specs + commands. Never report images that weren't generated.
• Don't re-produce what exists. An existing brief/logo/shot set skips its step; consume it instead.
• Facts stay verified. Model ID, script paths, and CLI flags come from the automation reference — never from memory.

Integration

Feeds this skill: skills/product-strategy/foundation-sprint (positioning and differentiators sharpen the brief in step 1 of routes A/D). Consumes its output: every skills/creative/* skill listed in the routing table — the plan tells each one what artifact it receives and what it must hand back.