/image-prompt

name: image-prompt
description: >-
  Use when the user wants to turn a short idea into a rich, production-grade
  image-generation prompt — posters, landing-page or UI mockups, ads, editorial
  layouts, photoreal scenes, game screenshots, logos, or illustrations. Builds ONE
  structured, copy-paste-ready prompt optimized for OpenAI GPT Image 2.6 by default, following
  proven layered-composition patterns (named zones, hex palette, real font references,
  mood/finish, IP-safety). Triggers on requests like "сделай крутой промт для постера",
  "нужна картинка/постер/лендинг для…", "generate a prompt for a landing hero",
  "prompt for an ad / logo / game screenshot".

image-prompt

Turn a one-line idea into a single, gold-quality image-generation prompt. Auto-first: infer everything sensible, output the prompt, let the user redirect in one line.

Workflow

1. **Read the idea.** Extract: subject/brand, any style hints, aspect/format, target model. Invent a fictional placeholder brand name if branding is implied but none given. If the user names a saved style ("in the <name> style", "use preset <name>") or asks to save one ("save this style as <name>"), handle it via **Style presets** below. If the user wants help choosing a look ("help me with the style", "what styles can you do", "в каких стилях можешь"), present the **Style menu** below. If the user supplies an existing image to transform, composite, restyle, localize, or place someone/something into — or wants a poster/ad built around a supplied person or product — that is an EDIT: use **Edit / remix mode** below.

2. **Pick the archetype** using `references/archetypes.md` (poster / landing-hero / product-ad / ui-mockup / photoreal-scene / game-screenshot / infographic / logo / illustration). Use the routing table. If ambiguous, choose the richest layout and note it.

3. **Set defaults** from the archetype: aspect → `size`, `quality`. (See the model adapter.)

4. **Expand into the 9-block skeleton** using `references/anatomy.md`. Skim the matching gold example(s) in `references/gold-examples.md` to calibrate depth and phrasing. Pull fonts + a cohesive hex palette from `references/fonts-palettes.md`. Fill every block: named zones with position/size%/tilt, literal copy in quotes, named fonts, hex palette, mood cluster, finish + quality tag. Photoreal/game ideas use the labeled-block variant.

5. **Format for the target model.** Default `gpt-image-2.6-sunburst` → follow `references/models/gpt-image.md` (labeled sections, text-in-quotes, CONSTRAINTS block, model/size/quality, SVG for logos, GIF for loops). If the user names another model and its file doesn't exist yet, say so and fall back to the rich natural-language style, then proceed.

6. **Output** (see format below): the finished prompt in its own fenced code block, then a one-line assumptions/knobs note underneath.

Auto-first — do NOT interrogate

Produce a full prompt on the first turn. Only ask **one** question if the idea is truly unworkable (e.g. no discernible subject at all). Everything else — archetype, aspect, palette, brand name, copy — you infer and surface in the assumptions line so the user can correct with a single sentence ("make it 9:16", "darker palette", "brand is NOVA").

Model handling

Default target: **gpt-image-2.6-sunburst** (Flare when the user wants speed). When the user names a model, load its adapter and follow it — the 9-block content stays the same; only the rendering (output shape, params/flags, size/ratio) changes.

| Model (match loosely, incl. RU) | Adapter | |---|---| | gpt-image-2.6 sunburst/flare (default), gpt-image-2.5, gpt-image-2, gpt-image-1.5, "openai", "гпт имедж" | `references/models/gpt-image.md` | | nano banana pro, gemini 3 pro image, "gemini", "нанобанана" | `references/models/gemini.md` | | midjourney, "mj", "миджорни" | `references/models/midjourney.md` | | flux, flux.2 | `references/models/flux.md` | | ideogram | `references/models/ideogram.md` | | recraft | `references/models/recraft.md` | | reve | `references/models/reve.md` | | universal / unknown tool | `references/models/universal.md` |

If the user names a model with no adapter yet, use `universal` and say so. **Pick by fit** when the user is unsure: text-heavy / infographic / factual → **gemini (Nano Banana Pro)** or gpt-image-2.6; brand / typography / vector / logo → **recraft**; pure aesthetics → **midjourney**; open-weight / local / hex control → **flux**; strong prompt-adherence + layout editing → **reve**; unknown tool or "just a great prompt" → **universal**.

Style presets (user-local, private)

Presets are saved **styles** (never content) the user has dialed in, stored per-user at `~/.claude/image-prompt/presets/<name>.md` — NOT in this repo, never committed, private to the machine. Create the directory on first save. Each user grows their own set.

• **Save** — on "save this style as <name>" / "запомни этот стиль как …": distill ONLY the

style layer (never subject, copy, or specific zones). Capture the **distinctive signatures that make the style non-generic** — palette + fonts alone flatten the output. Nail: **MEDIUM/technique** (watercolor vs flat vector vs risograph vs 3D render vs photo), **how and WHERE fills are painted** (flat vs mottled/grainy; applied globally vs confined to specific accent elements — apply a texture where the reference uses it, not everywhere), **LINE quality** (loose/varied vs clean/even), **COMPOSITION rhythm** (sparse/airy vs dense; organic vs grid; negative space), and an explicit **AVOID** list of anti-patterns that would drift toward a generic look. When deriving from reference images, name the 2-3 traits that separate this style from a generic version and encode them; if refs are available, sanity-check the intended output against them. A big AI-tell is the model's own default "polished illustration" look — soft digital watercolor gradien