GPT-Image-2 Prompt Library + Media Designer Agent
Two-piece skill:
scripts/search.py — thin HTTP client over a hosted corpus of 3,238 community-vetted prompts. BM25-ranked, tagged across 10 facets.agents/media-designer.md — agent profile that uses the search tool to turn a user brief into a paste-ready GPT-Image-2 prompt.
The tool finds candidates. The agent owns the judgement (which base, which slots to parameterise, which to keep literal, mood/palette fit).
When to Apply
Must use
- User wants a GPT-Image-2 / OpenAI image-generation prompt for a real production task (ad, poster, product shot, character sheet, UI mockup, portrait)
- User describes a brief and wants a polished prompt back, not just inspiration
- User is studying how top creators structure prompts and wants attributed examples
Skip
- User wants the image rendered — route to
ai-multimodal or ai-artist
- Task unrelated to image prompts
- User already has a finished prompt and just wants it run
Recommended Workflow
For any production prompt request:
1. Read agents/media-designer.md
2. Run the 6-step workflow it defines
3. Return the 4-block output (Base · Parameterised · Resolved · Rationale)
Hosted backend
Endpoint: https://gpt-image-2-prompts.goclawoffice.com
Thousands of community-vetted prompts indexed across 10 facets (subjects, styles, lighting, cameras, moods, palettes, compositions, mediums, techniques, usecases). Each record carries the prompt body, attribution, and a reference image. Rate-limited per IP — fair-use friendly, but please don't scrape.
CLI
search.py [query] [--shape SHAPE] [--has-image] [-n N] [--full] [--persist PATH]
# Free-text search (this is what the agent calls)
python scripts/search.py "luxury shoe ecommerce ad cream pastel" -n 5
# Narrow by shape when the brief is specific about format
python scripts/search.py "perfume bottle" --shape ecommerce -n 3
# Persist top hits as a markdown reference deck (with embedded images)
python scripts/search.py "neon ui" --persist plans/neon-refs.md
Filter knobs:
--shape — portrait | poster | ui | character | comparison | ecommerce | ad | thumbnail | infographic | comic--has-image — only records with a reference image-n N — top N (default 5)--full — don't truncate prompt body--persist PATH — write top results to a markdown file with embedded reference images
Output Anatomy
#1 bm25=-15.59 shape=ecommerce
id : z9q36mnc
title : Futuristic Bionic Super Shoe
author: @<creator>
tweet : https://x.com/<creator>/status/<tweet_id>
image : <reference image URL>
tags : subjects=product,fashion-item | styles=cinematic | cameras=low-angle |
moods=luxurious,intense,futuristic | palettes=gold-black |
techniques=parameterised-template
prompt:
Extreme futuristic {argument name="subject" default="cheetah bionic super shoe"} ...
Agent Profile
.claude/agents/media-designer.md defines the canonical workflow. Headline contents:
| Section |
Purpose |
| Mental model |
brief → diagnose → search → pick (mood-aware) → refactor → resolve → output |
| Step 1 — Diagnose |
Extract product, brand, shape, mood, palette, technique signals from the brief |
| Step 2 — Search |
Synthesise tokens, run search.py "<tokens>" -n 5 |
| Step 3 — Pick |
Mood-mismatch rejection table — pastel briefs reject moody/gritty/dark-amber, etc. |
| Step 4 — Refactor |
Replace product-specifics with {argument name="X" default="Y"} slots; keep mood/lighting/style words literal |
| Step 5 — Resolve |
Fill slots from user intent; default-fallback when ambiguous; never invent |
| Step 6 — Output |
4 blocks: Base (cite author + tweet) · Parameterised · Resolved · Rationale (≤80 words) |
1---2name: gpt-image-2-pro-max3description: Production prompt-engineering pipeline for GPT-Image-2 / OpenAI image generation. Pairs a 'media-designer' agent with a hosted searchable corpus of 3,238 community-vetted prompts, decomposed across 10 controlled vocabularies (subjects, styles, lighting, cameras, moods, palettes, compositions, mediums, techniques, usecases). Each record carries: full prompt body, twitter/X attribution link, downloaded reference image. Workflow: agent diagnoses the user brief → searches the corpus → picks a mood-aligned base → refactors the chosen prompt into a parameterised {argument} template → resolves arguments from user intent → returns the final paste-ready prompt with attribution + reference image. Use when the user wants a polished image-generation prompt for ads, posters, product shots, portraits, character sheets, UI mockups, infographics, exploded-view diagrams, or any other GPT-Image-2 / OpenAI image task.4---56# GPT-Image-2 Prompt Library + Media Designer Agent78Two-piece skill:9101. **`scripts/search.py`** — thin HTTP client over a hosted corpus of 3,238 community-vetted prompts. BM25-ranked, tagged across 10 facets.112. **`agents/media-designer.md`** — agent profile that *uses* the search tool to turn a user brief into a paste-ready GPT-Image-2 prompt.1213The tool finds candidates. The agent owns the judgement (which base, which slots to parameterise, which to keep literal, mood/palette fit).1415## When to Apply1617### Must use18- User wants a GPT-Image-2 / OpenAI image-generation prompt for a real production task (ad, poster, product shot, character sheet, UI mockup, portrait)19- User describes a brief and wants a polished prompt back, not just inspiration20- User is studying how top creators structure prompts and wants attributed examples2122### Skip23- User wants the image **rendered** — route to `ai-multimodal` or `ai-artist`24- Task unrelated to image prompts25- User already has a finished prompt and just wants it run2627## Recommended Workflow2829For any production prompt request:3031```321. Read agents/media-designer.md332. Run the 6-step workflow it defines343. Return the 4-block output (Base · Parameterised · Resolved · Rationale)35```3637## Hosted backend3839```40Endpoint: https://gpt-image-2-prompts.goclawoffice.com41```4243Thousands of community-vetted prompts indexed across 10 facets (subjects, styles, lighting, cameras, moods, palettes, compositions, mediums, techniques, usecases). Each record carries the prompt body, attribution, and a reference image. Rate-limited per IP — fair-use friendly, but please don't scrape.4445## CLI4647```48search.py [query] [--shape SHAPE] [--has-image] [-n N] [--full] [--persist PATH]49```5051```bash52# Free-text search (this is what the agent calls)53python scripts/search.py "luxury shoe ecommerce ad cream pastel" -n 55455# Narrow by shape when the brief is specific about format56python scripts/search.py "perfume bottle" --shape ecommerce -n 35758# Persist top hits as a markdown reference deck (with embedded images)59python scripts/search.py "neon ui" --persist plans/neon-refs.md60```6162Filter knobs:63- `--shape` — portrait | poster | ui | character | comparison | ecommerce | ad | thumbnail | infographic | comic64- `--has-image` — only records with a reference image65- `-n N` — top N (default 5)66- `--full` — don't truncate prompt body67- `--persist PATH` — write top results to a markdown file with embedded reference images6869## Output Anatomy7071```72#1 bm25=-15.59 shape=ecommerce73 id : z9q36mnc74 title : Futuristic Bionic Super Shoe75 author: @<creator>76 tweet : https://x.com/<creator>/status/<tweet_id>77 image : <reference image URL>78 tags : subjects=product,fashion-item | styles=cinematic | cameras=low-angle |79 moods=luxurious,intense,futuristic | palettes=gold-black |80 techniques=parameterised-template81 prompt:82 Extreme futuristic {argument name="subject" default="cheetah bionic super shoe"} ...83```8485## Agent Profile8687`.claude/agents/media-designer.md` defines the canonical workflow. Headline contents:8889| Section | Purpose |90|---|---|91| Mental model | `brief → diagnose → search → pick (mood-aware) → refactor → resolve → output` |92| Step 1 — Diagnose | Extract product, brand, shape, mood, palette, technique signals from the brief |93| Step 2 — Search | Synthesise tokens, run `search.py "<tokens>" -n 5` |94| Step 3 — Pick | Mood-mismatch rejection table — pastel briefs reject `moody/gritty/dark-amber`, etc. |95| Step 4 — Refactor | Replace product-specifics with `{argument name="X" default="Y"}` slots; keep mood/lighting/style words literal |96| Step 5 — Resolve | Fill slots from user intent; default-fallback when ambiguous; never invent |97| Step 6 — Output | 4 blocks: Base (cite author + tweet) · Parameterised · Resolved · Rationale (≤80 words) |
CongDon1207