This skill should be used when the user asks to "write a prompt for image generation", "design a generation prompt", "create a sprite prompt", "compose an OpenAI prompt", "structure a prompt for game art", or otherwise needs to construct prompts for the openai-creative-plugin. Covers the Identity Guide 5-line pattern, isolation suffix template for asset extraction, narrative-vs-keyword rules, and Do/Avoid prompting rules. Domain-specific prompt structures (sprites, environments, UI, icons, web) live in references/domain-structures.md.
How this skill is triggered — by the user, by Claude, or both
Slash command
/openai-creative-plugin:openai-promptsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Write effective prompts for GPT Image generation. The core principle: **describe the scene narratively, don't just list keywords.**
Write effective prompts for GPT Image generation. The core principle: describe the scene narratively, don't just list keywords.
[Subject] + [Composition] + [Action/State] + [Location/Context] + [Style] + [Technical specs]
Bad: knight pixel art game
Good: 2D pixel art knight character in idle stance, front-facing, blue tabard with gold lion emblem, silver plate armor, white background, 16x16 pixel grid, game sprite sheet style
For per-domain prompt structures (sprites, environments, characters, UI, icons, web sections), see references/domain-structures.md.
For series work — sprite sets, evolution-chain atlases, mockup screens, item packs — pin a byte-identical identity guide block at the top of every call in the series. The variant prompt then describes only what differs.
Five-line structure:
Identity: [IP name, what the world is, who plays]
Form: [3D/2D, geometry style, surface character, proportions]
Camera: [angle, lens, key/fill lighting setup]
Palette: [Name #HEX, Name #HEX, ...]
Materials: [textures, finishes, special-state effects]
Why it works: the model treats the guide as canonical context for what world this is in. The variant prompt only has to answer "what specific item this time?". Cross-call consistency comes from the guide being literally the same text every time, not from passing image references.
The Identity line is load-bearing. Give it a specific fictional project name — "Hollowfest Manor," "Crystal Caverns," "Tideglass Cove," whatever yours is. Real or invented; both work. Don't settle for category descriptors ("a Halloween merge-2 game") when an IP name is available. The model's prior for fictional-IP-cohesion activates when there's a name to anchor on — empirically a stronger consistency anchor than abstract world descriptors alone. Adapted from Dori Adar's original framing of the pattern.
When to use:
When NOT to use:
openai-workflows skill)Validated across a 27-image A/B/C experiment + 6-deliverable production pass.
⚠️ Provisional: Identity Guide validation was run on gemini-creative; re-validation on GPT Image pending.
For multi-image work, structure prompts in this labeled order — a skimmable template beats clever syntax. The Identity line comes first (load-bearing anchor; see above), then:
Identity: <IP/project name> — <asset family>
Scene: <background/context, or 'plain white background'>
Subject: <the thing, pose/state>
Details: <materials, palette, lighting>
Constraints: <what must NOT appear / must stay fixed>
Full guidance: references/openai-prompting-essentials.md
GPT Image handles text but demands precision:
quality: "medium" or "high" for small or dense text — "low" quality garbles itFull guidance: references/openai-prompting-essentials.md
Full guidance: references/openai-prompting-essentials.md
Do:
"Do not render color swatches, hex codes, or palette legends anywhere in the image." Validated at 5/6 (production pass) + 2/2 (#14b A/B) ≈ 88% suppression.Avoid:
When generating individual items destined to be cut out and used as standalone sprites/icons, append this hard-rules block verbatim:
Hard rules — must be followed exactly:
- Single object centered in frame, fully visible with breathing room from edges
- Pure white #FFFFFF uniform background, no gradient or vignette
- NO cast shadow, NO drop shadow, NO ground shadow — the item floats on the flat background
- NO color swatches, NO hex codes, NO palette legends, NO color charts anywhere
- NO UI elements, NO text, NO labels, NO decorations beyond what is described
Production-validated across 33 generations: ~100% shadow compliance, ~95% swatch suppression. The output is directly consumable by a chroma-key extraction pipeline. For white-ish subjects (ghosts, white skulls), fall back to semantic segmentation (rembg). See the openai-game-assets skill § Single-Isolated-Subject Extraction Pattern.
⚠️ Provisional: extraction validation rates inherited from gemini-creative; re-validation on GPT Image pending.
continue_editing with a targeted correction ("change the helmet visor to horizontal slats")generate_image — editing rarely fixes fundamental composition issuesreferences/domain-structures.md — Per-domain prompt structures: game sprites, game environments, character concepts, UI/mobile mockups, icon sets, web sections.references/openai-prompting-essentials.md — Distilled OpenAI cookbook best practices (anatomy, preserve-lists, ref indexing, text accuracy, photorealism, quality strategy, iteration).npx claudepluginhub 2901were/openai-creative-plugin --plugin openai-creative-pluginGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.