Image Prompt Skill Design

# image-prompt skill — design spec

**Date:** 2026-07-02
**Status:** approved, building v1

## Goal

A Claude Code skill: user throws a short idea, the skill auto-builds one
production-grade image-generation prompt following the proven layered-composition
pattern extracted from a corpus of high-quality reference prompts. Output is a
single clean, copy-paste-ready prompt (press `c` to copy) — no code, no API calls.

## Decisions (from brainstorming)

- **Output:** prompt text only. One clean copyable block + a short "assumptions/knobs"
  line so the user can redirect in one sentence. No runnable snippet, no auto-generation.
- **Default model:** `gpt-image-2`. Launch supports gpt-image-2 only.
- **Extensible:** adding a model later = adding one file under `references/models/`;
  the core builder is model-agnostic.
- **Auto-first:** infer archetype + defaults from the idea; do not interrogate.
  Ask at most one question only when genuinely ambiguous.

## The extracted pattern (9-block skeleton)

1. **Concept line** — format + medium + subject/brand + reference aesthetic + intended use.
2. **Canvas & background** — aspect ratio + background treatment (with hex) + global texture.
3. **Composition zones** — named regions with position + size% + tilt + content. The spine.
4. **Subject detail** — granular attributes; IP-safe "generic/original".
5. **Typography & text** — literal copy in quotes; per element: font style + named font
   reference + weight + tracking + case + color + effects + placement.
6. **Palette** — explicit hex list, each mapped to an element.
7. **Typography summary** — fonts mapped to roles.
8. **Mood** — adjective cluster + reference brands/eras.
9. **Finish / texture + quality tag** — post-processing + authority tag.

For gpt-image-2, also fold in a **CONSTRAINTS** block (no logos/trademarks/watermark,
original design, verbatim text) plus recommended `size` and `quality`.

## Two archetype families

- **Graphic design** (poster / landing-hero / editorial / product-ad / ui-mockup / logo /
  illustration): typography- and layout-heavy; all 9 blocks matter.
- **Photoreal / rendered scene** (candid photo / game-screenshot): labeled blocks
  (SCENE / SUBJECT / LIGHTING / CAMERA / [HUD] / TECHNICAL); palette/typography secondary.

## Architecture