Image generation

Image Generation

Convert a fuzzy image request into prompt specs precise enough that any competent image model — and specifically fal-ai/nano-banana-pro on FAL.ai, the one model this repo's automation uses — renders what the user actually pictured. The prompt spec is the deliverable; generated files are the bonus when the automation is configured. Prime rule: never generate from a vague prompt. A weak prompt costs a generation credit and teaches you nothing; a sharp spec is reusable forever.

When to use / when not to

Use for general-purpose images: hero images, illustrations, thumbnails, infographics, textures, concept art, any one-off visual. Hand off instead when the request is really:

• Product shots for commerce/marketing → skills/creative/product-photography
• Platform-sized social content → skills/creative/social-graphics
• Logos, icons, patterns, brand marks → skills/creative/brand-asset
• No visual direction exists yet and consistency matters → run skills/creative/creative-strategist first and consume its style block
• A whole multi-asset campaign → skills/creative/orchestrator routes it

Intake

Ask in one batch, only what's missing:

1. Subject — what exactly is in the image? (object, scene, people, text?)
2. Use — where does it go? (this decides aspect ratio and resolution)
3. Style anchor — existing style guide / creative-strategist style block, reference images, or 3 adjectives for the mood?
4. Automation — is FAL_API_KEY set so you can generate, or deliver specs only?

Infer, don't ask: aspect ratio from the stated destination (thumbnail → 16:9, story → 9:16, feed → 1:1 or 4:5); resolution from stakes (2K default, 4K hero/print, 1K drafts). Don't stall: if subject and use are clear, state your assumptions for the rest and proceed.

Workflow

1. Check for a style block. If the user has run skills/creative/creative-strategist, reuse its style block verbatim in every prompt. If not, and this is a one-off, derive a minimal one (style + palette + mood) from context and label it an assumption.

2. Draft the prompt with the 5-part formula. Every prompt contains, in order: subject (the main thing, specific: "a luxury leather watch with gold accents", never "a watch") → descriptive detail (materials, colors, text to render) → style ("professional product photography", "flat modern illustration", "photorealistic") → technical treatment (lighting, composition, focus, "4K, highly detailed") → mood ("warm and inviting"). Add negative constraints as plain prohibitions ("no text overlay", "no watermark", "no busy background").

3. Decide variant strategy. If it's a throwaway draft → 1 image at 1K. If it's a keeper asset → propose 3–4 genuinely distinct prompt specs (different composition or style angle, not synonym swaps), generate 1 of each or num_images 3–4 of the winner, and have the user pick. Never present one option for a hero asset.

4. Generate or deliver. If FAL_API_KEY (or FAL_KEY) is set:

   python docs/creative_cli.py custom \
     --category "<category>" --name "<asset-name>" \
     --prompt "<full prompt>" \
     --aspect-ratio 16:9 --resolution 2K --num-images 3
   

   Use --web-search only when the image must reflect current real-world data (e.g. "2026 trends infographic"). If no key is set, deliver the prompt-spec blocks and the exact command the user can run later — do not fake results or claim images were generated.

5. Review and iterate. Compare output to the spec's mood/style/negative constraints. Fix misses by editing the prompt (more specific subject, explicit lighting, stronger prohibitions) — there are no other model knobs. One iteration round maximum before checking in with the user.

Read references/automation.md before writing any generation command — it has the full, verified parameter set, CLI flags, Python helpers, and output layout.

Required output format

Deliver one block per asset (this exact structure, whether or not you generate):

## Prompt Spec — [asset name]
- **Use / destination:** [where it will live]
- **Subject:** [the main thing, concretely]
- **Style:** [artistic treatment]
- **Composition:** [framing — centered / rule of thirds / negative space / overhead]
- **Lighting:** [studio / natural golden hour / dramatic rim / high-key]
- **Mood:** [2–3 adjectives]
- **Aspect ratio:** [one of 21:9 16:9 3:2 4:3 5:4 1:1 4:5 3:4 2:3 9:16] · **Resolution:** [1K/2K/4K] · **Format:** [png/jpeg/webp]
- **Negative constraints:** [what must NOT appear]
- **Prompt (final, paste-ready):**
  > [single flowing prompt assembled from the fields above]
- **Variants:** [n] — [what varies between them]
- **File naming:** assets/<category>/<asset-name>/<asset_name>_<n>_<timestamp>.png (automation default)
- **Generate with:** `python docs/creative_cli.py custom --category ... --name ... --prompt "..." --aspect-ratio ... --resolution ... --num-images n`

If images were generated, append the saved paths under Results and say which variant you recommend and why (one line).

Quality bar

Before delivering, verify every spec:

• Subject is concrete enough that two strangers would picture the same image
• Style, lighting, composition, and mood are each explicitly stated — none left to model default
• Aspect ratio matches the stated destination (no 1:1 "because default")
• Negative constraints listed (at minimum: no watermark; state text policy explicitly — models mangle long text)
• No invented parameters — only the ones in references/automation.md; the model is always fal-ai/nano-banana-pro
• Keeper assets got 3+ distinct variants, not one take
• If automation wasn't run, the deliverable says so and includes runnable commands

Hard don'ts: never claim an image was generated when it wasn't; never put brand color hex codes in quotes and hope — name the colors in words too ("deep navy #004E89"); never render paragraphs of text inside an image.

Integration

• skills/creative/creative-strategist → feeds this skill its style block (paste into every prompt's style/mood/lighting fields).
• skills/creative/product-photography, skills/creative/social-graphics, skills/creative/brand-asset → specialized front-ends that produce their own prompt specs and use the same automation; route there when the request fits.
• skills/creative/product-video and skills/creative/remotion-script-writer → consume generated stills as source frames / assets.
• references/automation.md — full FAL.ai client, CLI, and Python-helper reference. Read it whenever you're about to run a generation command.