# Image Gen

> Image Gen

- Skill: `trac3r00/image-gen` (Agent Skill)
- Install (CLI): `npx skillmds@latest add trac3r00/image-gen`
- Raw SKILL.md: https://api.skillmd.com/api/skills/trac3r00/image-gen/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: trac3r00 (https://skillmd.com/u/trac3r00)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/trac3r00/image-gen

---


# Image Gen

The first generation should be close. Provider choice and prompt structure
decide that, not luck.

## Provider routing

| Job | Provider | Why |
|---|---|---|
| Text in image (posters, UI mockups, signage) | gpt-image (gpt-image-2) | Best-in-class verbatim text rendering |
| Design fidelity, brand-adjacent work | gpt-image | Follows structured art direction |
| Fast iteration, photoreal people | grok (via groken cloud) | Speed, identity realism |
| Textures, tiles, sprites | gpt-image low quality first, then high | Cheap drafts before the final render |

Check what's installed before promising: the harness's `generate_image`
tool for gpt-image; the groken bridge for grok. Never fake a provider that
isn't there — say what's missing.

## Prompt structure (the part people skip)

1. **Subject** — one clear subject, named concretely ("a ceramic pour-over
   coffee dripper" not "coffee thing").
2. **Composition** — where the subject sits, camera angle, framing
   ("centered, slight low angle, product shot on seamless").
3. **Style** — the visual language ("editorial product photography",
   "flat vector illustration", "35mm film grain").
4. **Lighting** — the single biggest quality lever ("soft key from upper
   left, subtle rim light", "overcast window light").
5. **Negatives** — what to avoid ("no text, no watermark, no hands").
6. **Verbatim text** — if text appears, put it in quotes exactly
   (`the words "SALE ENDS FRIDAY"`) and keep it short.

## Iteration discipline

Draft cheap (low quality), pick a direction, then render the final at high
quality. Read the model's revised_prompt when available — it tells you what
it actually heard. One strong revision beats five blind regenerations.

## Pairs with

`design` (the taste standard the image is judged against),
`higgsfield-generate` (heavier video/3D/audio generation when installed).