Stats
Actions
Tags
This skill should be used when the user asks to "generate game assets", "create pixel art", "make a sprite", "design a tileset", "generate a tier evolution atlas", "create a merge-game asset pack", "design game items", or otherwise needs to produce pixel art / game art via the openai-creative-plugin. Covers the Identity Guide pattern (with pixel-art-specific alternative), tier-atlas generation, isolated-subject extraction patterns, and guide overreach workarounds. Palette/resolution tables and sprite/tileset workflows live in references/.
How this skill is triggered — by the user, by Claude, or both
Slash command
/openai-creative-plugin:openai-game-assetsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Generate consistent pixel art game assets — sprites, tilesets, environments, and UI. Works best with a defined style anchor before generating anything.
Generate consistent pixel art game assets — sprites, tilesets, environments, and UI. Works best with a defined style anchor before generating anything.
model: "gpt-image-1.5" + background: "transparent".
The default model (gpt-image-2) REJECTS transparent backgrounds — switch models before generating.gpt-image-2.Before generating a single image, establish a style anchor — a block of text you'll pin byte-identical at the top of every prompt in the series. Two validated formats; pick the one that fits your project.
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]
The Identity line is the load-bearing anchor. Give it a specific fictional project name ("Hollowfest Manor," "Crystal Caverns," whatever yours is) — the IP name activates the model's prior for fictional-world cohesion. Empirically faster convergence than abstract category descriptors ("a Halloween merge-2 game") alone. See openai-prompts skill § "The Identity Guide Pattern" for the full rationale.
Use this for projects that mix asset types (3D items + UI screens), or when one guide needs to anchor cross-modality output. Validated across a 27-image A/B/C experiment + 6-deliverable production pass.
Resolution: [32x32 characters, 16x16 tiles]
Palette: [PICO-8 / 16 colors]
Outlines: [1px black / colored / none]
Shading: [flat / 2-tone cel / gradient + dithering]
Animation: [80ms base, walk 6f, jump 4f, attack 5f]
Proportions: [2.5 heads tall, hero 3x tile height]
Use this when the project is pure pixel-art / sprite-based and animation timing / outline rules matter at the prompt level — the per-pixel constraints don't fit cleanly into the 5-line Identity Guide. This pixel-art bible is the standalone style prefix for sprites that aren't part of an evolution chain or unified-IP series.
Rule for either format: every asset must look like it was made by the same person on the same day. The guide enforces this without visual references — but only if the guide text is byte-identical across calls. Re-describing in different words causes drift (validated by Mode-B paraphrase control: "Lighthouse" came out as a tavern).
⚠️ Provisional: paraphrase-drift validation was run on gemini-creative; re-validation on GPT Image pending.
For resolution tables, named palettes, and quick-start style prefixes, see references/palette-and-resolution.md; for character, tileset, and UI step-by-step workflows, see references/sprite-workflows.md.
For merge-2 / progression games, generate a multi-tier evolution chain as a single horizontal atlas in one call instead of N separate generations:
[GUIDE BLOCK — pasted verbatim]
Compose a horizontal evolution-chain atlas: [N] [item] merge-tier items
arranged left-to-right in a single row on a neutral cream-white background
with subtle vertical divider lines between each tier. Same camera angle,
same lighting, same proportions on every tier — the items must read as the
same item family at progressing power levels.
Tier 1 — [Name]: [short description]
Tier 2 — [Name]: [short description]
Tier 3 — [Name]: [short description]
... (etc)
Do not render color swatches, hex codes, or palette legends anywhere in the image.
Critical: atlas-shaped variants (dense labels + multi-tier composition) historically trigger a palette-legend rendering bug where the model renders the guide's Palette line as visible swatches inside the image. The negative-instruction suffix above is the validated mitigation — 5/6 (production pass) + 2/2 (#14b A/B) ≈ 88% suppression on gemini-creative. Append it verbatim.
⚠️ Provisional: palette-legend suppression rate inherited from gemini-creative validation; re-validation on GPT Image pending.
Aspect ratio matters. Pick one that natively fits a 1×N row:
Portrait aspects (9:16, 4:5) will silently reshape into a 2-row grid.
Known failure mode (low-prevalence): the model occasionally renders TWO rows of N tiles instead of the requested single row, with mildly garbled labels (observed in #14b treatment trial). Re-roll if this happens — no prompt-level fix has been validated.
For generating individual sprites destined for chroma-key extraction → transparent PNG pipeline:
[GUIDE BLOCK — pasted verbatim]
[Variant description — one focused subject only, no atlas/multi-element framing]
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 in a single asset pack: ~100% shadow compliance, ~95% swatch suppression. The output is directly consumable by a chroma-key-on-white extraction pipeline for most subjects. For white-ish or pale subjects (ghosts, white skulls, cobweb) chroma-key won't work — fall back to semantic segmentation (rembg) per asset. Default chroma-key, opt-in rembg per-asset.
⚠️ Provisional: extraction-pipeline validation rates inherited from gemini-creative; re-validation on GPT Image pending.
Alternative for transparency: with gpt-image-1.5 you can request background: "transparent" directly instead of post-processing chroma-key. Use gpt-image-1.5 + transparent background for any isolated subject that needs a true alpha channel.
Transparent extraction phrasing (gpt-image-1.5): Request "transparent background (RGBA PNG), crisp silhouette, no halos or fringing." Optionally add: "subtle, realistic contact shadow in the alpha (no hard cut line)." For extractions from scenes: "remove the background only — do not restyle the subject." (Source: OpenAI cookbook)
Full guidance: the openai-prompts skill → references/openai-prompting-essentials.md
Once a Identity Guide is established, the model can over-apply its visual motifs to items that shouldn't inherit them. Observed in the 2026-05-12 Hollowfest Manor pack:
Materials: hand-stitched felt)Workaround: for baseline items that shouldn't carry the decorative motifs, add explicit negation in the variant: "just the plain [item] form, no extra decorations, no felt patches, no [theme-specific] embellishments." Single-shot observation in one production pass; flag if you see it in your own work so we can refine the workaround.
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.