Image prompt

<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	Always at step 3 — intent → model → which model-file to load
references/prompt-formula.md	When building any prompt — 6-part formula + per-part vocabularies
references/lighting-vocabulary.md	When picking lighting hints — portrait / scene / quality-of-light dictionaries
references/camera-vocabulary.md	When the image should look photographic — lens/sensor/quality-tag dictionary
references/editing-prompting.md	Mode edit or multi-ref — preserve/change grammar, identity locks, weighted refs
references/text-in-image.md	Mode text-heavy — per-model rules for legible text + multilingual
references/models/midjourney.md	Midjourney V8.1 (v7 legacy flags) — --sref, --oref, --raw, --ar, --s, --c, --no, --p, --w
references/models/flux.md	Flux 2 Pro/Dev, 1.1 Pro Ultra (Raw), Kontext, Schnell, Krea
references/models/google.md	Nano Banana Pro / 2 / 2 Lite (Gemini image family) + the Imagen 4 shutdown
references/models/openai.md	gpt-image-2 (DALL-E 3 retirement note)
references/models/ideogram-recraft.md	Ideogram 3 Flash/Turbo/Default/Quality, Ideogram 4 (open weights, JSON prompting) + Recraft V3 (SVG)
references/models/bytedance-seedream.md	Seedream 5.0 Pro / Lite (weighted multi-ref, layered PNG output)
references/models/open-source.md	SD 3.5 + SDXL legacy + Qwen-Image 2.0 + HiDream-O1
references/execute.md	--execute mode — env var matrix, provider availability check, cost preview, troubleshooting, fall-back behaviour

EXAMPLES

See examples/before-after.md — calibration pairs covering portrait, product, scene, abstract, illustration, text-in-image (Ideogram 3), edit (Flux Kontext), multi-reference composite (Seedream 5.0), open-weights (Qwen-Image).

CONSTRAINTS

• Don't name real people. Use role descriptors ("a confident business person") not real names.
• Don't promise reliable text > 1 short phrase outside text-tier models. Reliable text: Ideogram 3 Quality, Nano Banana Pro, Nano Banana 2, gpt-image-2, Qwen-Image. Avoid > 5 words in Midjourney / Flux 2 Pro (≈60%); avoid entirely in SD 3.5 / SDXL. For exact typography: render in a design tool, not the generator.
• For character consistency across multiple images — use a model that supports identity locks: Midjourney V8.1 (--oref), Flux Kontext, Nano Banana Pro (up to 5 people), gpt-image-2 (16 refs), Seedream 5.0 (Character weight 1.0).
• Don't re-describe the ref's appearance in the prompt. When --reference is attached as Character, the prompt should describe wardrobe / action / expression / environment — NOT face / hair / body. Re-describing overrides the ref and causes drift.
• Don't mix Midjourney -- flags into NL-only models. Nano Banana / gpt-image-2 / Flux NL prompts ignore (or break on) --ar, --s, --style raw, etc. Use API params or NL phrasing instead.
• SD 3.5 weight syntax (word:1.3) is a no-op. Despite accepting the syntax, SD 3.5 ignores weights. Use keyword priority order. Weights DO work on SDXL / SD 1.5.
• One subject per prompt unless explicitly multi-subject. The model can't render "a cat, a dog, a horse, and a fox" cleanly.
• Specific lighting beats abstract. "Soft directional key light from upper left" > "good lighting".
• Don't bury the subject. First 12-15 words anchor the model.
• Kontext deviation: in edit mode with Flux Kontext / Nano Banana Pro / gpt-image-2, the prompt is JUST the change instruction — don't restate subject / setting / style from the source image.
• --execute is opt-in. Default flow stays prompt-only. Only run the API when the user passes --execute.
• Never print API keys. Not in output, not in errors, not in fall-back text. Mask if you must reference them ("set $OPENAI_API_KEY", not "key starts with sk-...").
• Confirm cost. Anything above $0.10 estimated must hit interactive Y/N (handled by common/runners/cost.py). Bypass only when user passes --yes.
• Output dir is ./generated/image/ by default. Don't write outside it without explicit --output.
• API failure → fall back gracefully. Save prompt to ./generated/image/<timestamp>-prompt-only.txt with a one-line reason. Skill stays useful.

INVOCATION HINTS

When the user says any of:

• "generate / write / make me a prompt for an image / cover / illustration / artwork"
• "Midjourney / Flux / Imagen / Nano Banana / gpt-image-2 / Ideogram / Seedream / Recraft / Qwen-Image / SD prompt for..."
• "edit this image", "change the dress color, keep the face", "preserve identity"
• "character consistency across covers / chapters"
• "multi-reference", "composite from these images"
• "poster / book cover / signage with legible text"
• "cover image / hero image / thumbnail / illustration prompt"
• "product shot / portrait / scene prompt"
• "improve this image prompt"
• "execute the prompt", "actually generate", "fire the gen", "use my OpenAI / Flux / Imagen key"
• "save the image", "render the asset"

RU triggers (use the skill when the user writes any of):

• «промпт для Midjourney / Flux / Imagen / Nano Banana / gpt-image-2 / Ideogram / Seedream / Recraft / Qwen / HiDream / SD»
• «обложка для статьи / поста / лонгрида»
• «отредактируй картинку», «поменяй цвет платья, оставь лицо», «сохрани идентичность»
• «единый персонаж на всех обложках», «character consistency»
• «постер с текстом», «обложка книги с подзаголовком»
• «hero-картинка для лендинга»
• «улучшить промпт для изображения»
• «multi-reference композит», «комбинация рефов»
• «выполни промпт», «сгенерируй через API», «вызови модель», «сделай реально»
• «используй мой OpenAI / Imagen / Flux ключ», «сохрани картинку»

The prompt itself is usually written in English (most models parse EN best). Only when the user explicitly asks for an RU-language prompt should the body be RU. RU terminology mapping for lighting + camera vocabulary lives in references/lighting-vocabulary.md (section RU терминология).

For multilingual text rendering INSIDE the image (Chinese, Japanese, mixed scripts), use Qwen-Image 2.0 — see references/text-in-image.md.

Use this skill. For video — use video-prompt (different vocabulary, has temporal flow + camera movement, not single-frame lighting).