/image-prompt

name: image-prompt
description: "Write prompts for 14+ frontier AI image generators (Midjourney V8.1, Flux 2/Kontext, Nano Banana Pro/2, gpt-image-2, Ideogram 4/3, Recraft V3, Seedream 5, Qwen, HiDream, Krea, SDXL). Modes: T2I, edit, multi-ref, text-in-image. Use when: 'prompt for an image', 'Midjourney prompt', 'edit with Kontext', 'character consistency', 'poster with text'."

license: MIT
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Grep
  - Glob

<objective> Generate a tight, model-aware prompt for AI image generation. Output: one well-formed prompt string (optionally + a negativePrompt, reference list, or edit instruction). This skill does NOT call the image model — it produces the text you paste into Midjourney / Flux / Imagen / Nano Banana / gpt-image-2 / Ideogram / Recraft / Seedream / Qwen-Image / HiDream / Krea / SD.

Use when the user wants an image for a specific scene, post cover, product mock, portrait, hero illustration, abstract background, OR wants to edit an existing image with character preserved, OR wants a multi-reference composite (character + style + palette), OR wants legible text inside an image. The skill picks the right model + mode + vocabulary that produces sharp, coherent output instead of generic "AI image".

This skill does NOT:

• generate the image itself (that's the model)
• design layouts or multi-frame compositions (use a design tool)
• write video prompts (use `video-prompt`)

</objective>

ROLE

Read the request → identify subject + intent (generate / edit / multi-ref / text-heavy) → pick target model from `references/model-picker.md` → assemble the prompt using the 6-part formula plus the conditional 7th block (references) → add lighting + camera + texture hints → return the prompt + optional negative or reference list.

PIPELINE

1. **Clarify if needed.** If the user gave only a topic ("cover for a post about cold-emails"), pick a sensible default subject ("close-up of an empty mailbox at dawn") and check before committing. If they gave a full scene, skip.

2. **Mode select.** Pick one:

• `t2i` — text-to-image (default).
• `edit` — modify an existing image; preserve identity / lighting / pose, change one thing.
• `multi-ref` — compose from multiple reference images (character + product + style).
• `text-heavy` — legible text is the subject (poster, book cover, signage).

3. **Pick model.** Default by intent — see `references/model-picker.md`:

• Editorial / fashion / "vibes" → Midjourney V8.1
• Photoreal portrait / product → Flux 2 Pro or Nano Banana Pro
• Text-heavy → Ideogram 3 Quality or Nano Banana Pro
• Edit → Flux Kontext or Nano Banana Pro or gpt-image-2
• Multi-ref composite → Seedream 5.0 (weighted roles, layered output) or Flux 2 Pro or gpt-image-2
• Self-host / open-weights → Flux 2 [dev], SD 3.5, Qwen-Image 2.0 (CJK), HiDream-O1
• Cheap iteration → Flux Schnell / Nano Banana 2 Lite / Ideogram 3 Turbo

4. **Build the prompt** — see `references/prompt-formula.md`:

   {subject + action} + {setting} + {style} + {lighting} + {camera/lens} + {texture/realism}

When mode is `edit` or `multi-ref`, the 7th conditional block fires — see `references/editing-prompting.md`.

5. **Load model-specific syntax** from `references/models/<vendor>.md` and apply (flags / NL phrasing / weighted refs / preserve-change grammar).

6. **Add negative prompt** if useful (mainly SDXL / Flux). Standard set: `text, watermark, logo, distorted anatomy, extra fingers, blurry, low resolution`.

7. **Output.** Return:

• The prompt as one fence-block (paste-ready)
• Optional negative as a second fence-block
• For multi-ref: an annotated list of refs with roles + weights
• 1-line note: which model + mode + key conventions applied
• If `--variants N` requested — N alternatives with different style / lighting / camera

8. **(Optional) Execute via API.** If `--execute` was passed AND the env var for the chosen model is set, also run `python3 scripts/run.py --model <model> --prompt-file <generated.txt>`. This calls the vendor API and saves a real PNG to `./generated/image/`. On any failure, fall back to prompt-only and print the reason. See `references/execute.md`.

MODES

• `image-prompt <topic-or-scene>` — generate default prompt (intent-routed model)
• `image-prompt <scene> --model <name>` — target a specific model. Valid: `midjourney-v8`, `flux-2-pro`, `flux-2-dev`, `flux-1-1-pro-ultra`, `flux-kontext`, `flux-schnell`, `flux-krea`, `nano-banana-pro`, `nano-banana-2`, `nano-banana-2-lite`, `gpt-image-2`, `ideogram-3`, `recraft-v3`, `seedream-5`, `qwen-image`, `hidream-o1`, `krea-1`, `sd-3-5`, `sdxl`
• `image-prompt <scene> --style <style>` — force a style (`photorealistic`, `editorial`, `3d-render`, `illustration`, `product-shot`, `cinematic`, `minimalist`, `no-ai-look`)
• `image-prompt <scene> --edit` — edit mode; expects a source image (URL or path). Generates preserve/change instruction.
• `image-prompt <scene> --reference <path-or-url>[@<role>:<weight>]` — attach a reference. Repeatable. Roles: `character`, `style`, `palette`, `layout`. Weights 0-1. Triggers multi-ref mode.
• `image-prompt <scene> --variants 3` — 3 alternatives with different style or lighting
• `image-prompt <scene> --improve` — user provides a weak prompt + the bad output description; skill rewrites
• `image-prompt <scene> --execute` — also call the API if the env var for `--model` is set; save PNG to `./generated/image/`
• `image-prompt <scene> --execute --output <dir>` — custom output dir
• `image-prompt <scene> --execute --yes` — skip cost confirmation
• `image-prompt --check --model <slug>` — verify env + connectivity, no generation
• `image-prompt --list-providers` — list executable providers given current env (image modality)

REFERENCES (load on demand)

| File | When to load | |---|---| | [references/model-picker.md](references/model-picker.md) | Always at step 3 — intent → model → which model-file to l