gpt-image-prompt-craft

---
name: gpt-image-prompt-craft
description: Use when turning a rough image idea into a production-grade GPT Image 2 prompt — applies craft rules (exact text in quotes, canvas-first layout, JSON render schemas, material/lighting controls) and a 162-prompt gallery.
source: https://github.com/wuyoscar/GPT-Image2-Skill
generated: 2026-07-02T18:41:50.336Z
category: tool
audience: creators
---

## When to use

- Enhancing a weak or vague image prompt into a precise, render-ready spec for GPT Image 2 / gpt-image-2
- Generating posters, typography, dense Chinese text, UI mockups, infographics, or research/data figures that need exact labels and layout
- Writing JSON/config-style prompts for premium product or food renders with material, lighting, and motion subsystems
- Editing or inpainting reference images while preserving identity, layout, text, and brand invariants
- Routing to a matching pattern from the 162-prompt Reference Gallery before writing a prompt from scratch

## Key concepts

### Canvas-first structure

State aspect ratio and layout before the subject. If structure matters, describe it first — otherwise the model spends its detail budget on the object and improvises the layout.

### Exact text in quotes

GPT Image 2 renders typography well only when literal copy is explicit. Wrap every displayed string in "…", keep user Chinese verbatim, separate blocks with /, and mark whether text must be crisp/legible vs decorative.

### JSON render schema

For complex multi-system renders (product hero, food, motion), use a structured config prompt with GLOBAL_SETTINGS / ENVIRONMENT / CORE_ASSETS / MOTION / OUTPUT slots and render_flags — concrete visual constraints, not vague praise.

### Material / lighting / palette are separate controls

Don't compress into 'premium'. Split into concrete materials (brushed steel, condensation), lighting (softbox, rim light, side light), and a bounded palette (muted teal/rust/bone).

### Scene density beats adjectives

Name 5-12 concrete nouns plus 2-4 material/lighting constraints instead of stacking empty adjectives like 'stunning' or 'high quality'. Capture context (RAW iPhone photo, 28mm lens feel) unlocks photorealism.

### Edit invariants

For images.edit, name the transformation first and explicitly preserve identity/layout/text/position. Identify multi-references by index and role ('Image 1: product, Image 2: style') and repeat invariants each iteration.

## API reference

```
gpt-image -p "PROMPT" [-f OUT] [-i REF...] [-m MASK] [--size --quality]
```

Run generation via the packaged CLI. No -i = text-to-image; one+ -i = reference edit; -i plus -m = inpaint with mask. Reads OPENAI_API_KEY; high quality for final assets, low/medium for drafts.

```
# text-to-image
gpt-image -p "Design a 3:4 vertical poster…" --size portrait --quality high
# reference edit
gpt-image -p "change only the season to winter; keep position" -i scene.png
# inpaint
gpt-image -p "remove the logo" -i poster.png -m mask.png
```

```
Exact text: wrap every displayed string in quotes
```

Make typography reliable by stating literal copy instead of describing it.

```
Weak:  Create a tea poster with the brand name and promo copy.
Strong: Design a 3:4 vertical poster. Accurately display the exact Chinese copy: "山川茶事" / "冷泡系列" / "中杯 16 元" / "大杯 19 元". Crisp, legible, no garbled characters.
```

```
JSON render config schema (premium product/food)
```

Structured config prompt for outputs with many interacting systems — environment, materials, lighting, motion, render goals.

```
/* PRODUCT_RENDER_CONFIG: Hero
   AESTHETIC: Premium Commercial Photography */
{
  "GLOBAL_SETTINGS": { "aspect_ratio": "2:3 vertical", "style": "hyper-realistic commercial photography", "render_flags": ["8K_UHD", "sharp_foreground", "editorial_finish"] },
  "ENVIRONMENT": { "background": "warm gradient studio", "lighting": "directional softbox, glossy highlights", "atmosphere": ["floating particles", "cinematic bokeh"] },
  "CORE_ASSETS": { "primary_subject": "hero product", "materials": ["brushed metal", "condensation"], "composition": "diagonal zero-gravity arrangement" },
  "OUTPUT": { "mood": "premium, editorial", "avoid": ["cheap banner", "plastic CGI", "fake brand logos"] }
}
```

```
Edit prompt: name the change + preserve invariants
```

Surgical edits that transform one thing while explicitly keeping the rest readable and in place.

```
Make it a winter evening with heavy snowfall, snow dusted on the board and pieces, breath vapor, cold blue-grey lighting — chess position still clearly readable, pieces in the original positions.
```

## Gotchas

- Requires OPENAI_API_KEY and the gpt-image CLI (Python 3.11+, uv/uvx); successful API calls bill the user's OpenAI account.
- Don't reimplement image-gen code for normal requests — call the packaged gpt-image CLI or scripts/generate.py.
- Search the bundled gallery/craft references before writing from scratch; load the smallest useful slice, never all category files.
- Use high quality for final assets, Chinese text, posters, diagrams, and UI; low/medium for drafts and broad exploration.
- Real-person likeness edits often fail at the API moderation layer — surface the API error verbatim rather than working around it.
- Keep negation/avoid-lines short and targeted; too many negatives dominate the prompt.
- Pick one dominant capture frame for photorealism — conflicting camera specs degrade the result.

---
Generated by SkillMake from https://github.com/wuyoscar/GPT-Image2-Skill on 2026-07-02T18:41:50.336Z.
Verify against source before relying on details.

File: ~/.claude/skills/gpt-image-prompt-craft/SKILL.md