gamedev-skills/awesome-gamedev-agent-skills / skills/disciplines/create-game-assets
create-game-assets
VerifiedPlans, generates or sources, and validates consistent game art: style guides, sprites, pixel art, tilesets, icons and textures, with QA and licence records. Use when making or AI-generating game art, fixing fake transparency, or checking Steam AI disclosure. Not for engine import (`godot-tilemap`, `godot-animation`, `unity-tilemap-2d`), UI layout (`game-ui-ux`) or sound (`audio-design`).
Create Game Assets
Turn a game's visual intent into a consistent, engine-ready asset set. Treat image generation as one production tool inside a controlled pipeline, never as proof that an asset is shippable.
When to use
- Defining or refreshing an art direction, or writing a style guide other contributors follow.
- Producing a family of sprites, tiles, icons, backgrounds or textures that must match.
- Generating art with an image tool, or sourcing licensed packs, and making the result usable.
- Auditing what was generated or licensed before release.
Not this skill: importing into the engine (godot-tilemap, unity-tilemap-2d,
pixijs-rendering, threejs-gltf-loading), animating the sprites (godot-animation,
unity-animation), atlas and texture-memory budgets at runtime (performance-optimization).
Core workflow
- Inspect before inventing. Find existing screenshots, concept art, sprites, models, fonts, import settings, camera framing, target resolution, and naming conventions. Preserve a coherent existing direction unless the user asks for a redesign.
- Lock the technical frame. Record engine, 2D/3D, camera/view, native display size, asset
dimensions, world scale, transparency, palette, filtering, animation frames, texture budget,
and target platforms. Use
assets/art-direction-brief.mdas a copyable brief. - Name the visual system. Define shape language, silhouette priorities, value structure, palette roles, materials, lighting, detail density, edge treatment, and motion character. Use concrete visual properties; do not substitute a living artist's name for an art direction.
- Make an asset manifest. Copy
assets/asset-manifest.json, then list every required asset, state, variant, size, pivot, collision role, source, license, and approval status. Separate production assets from disposable greybox placeholders. - Approve one visual target. Create or select a representative hero asset or small style board before producing a full set. Judge it at actual game scale and against a gameplay background. If the user delegated the choice, pick the strongest viable direction and record the decision instead of blocking.
- Produce related assets as families. Reuse the approved target as an edit/reference input. Keep palette, view, proportions, lighting, outline, and texture density invariant. Generate small coherent batches; avoid unrelated one-off prompts that drift.
- Normalize deterministically. Crop, size, anchor, slice, name, compress, and check alpha with ordinary image/DCC tools. Never trust generated grids, transparency, seams, pivots, topology, or dimensions without inspection. Use the bundled scripts for raster QA and contact sheets.
- Import with engine-native settings. Set filtering, mipmaps, pixels-per-unit/world scale, color space, compression, sprite slicing, texture types, materials, and collision deliberately. Read the relevant engine skill before editing engine files.
- Validate in context. Inspect a contact sheet and the actual game at native resolution. Check silhouette, scale, animation stability, seams, legibility, palette, collision fit, memory, and compression artifacts. Iterate on the source asset, not only on runtime compensations.
- Record provenance. Keep the source URL/tool, license or generation note, edit history, and restrictions beside the manifest. Preserve embedded provenance metadata when the pipeline can.
- Finish with the checks. Run
scripts/asset_report.pyover every file in the batch with the manifest's size, alpha and colour limits and--json; it must exit 0. Build a contact sheet withscripts/build_preview_sheet.pyand look at it. Confirm every shipped manifest entry has a source, licence andtechnical_checks_passed: true. Then ask the user to view the assets in the running game at native resolution and report anything that does not read.
Choose the production path
| Need | Default path |
|---|---|
| Existing asset needs a controlled change | Edit the original/reference; state what must remain unchanged |
| New 2D visual family | Approve seed → generate/source family → normalize → preview → import |
| True pixel art | Use generated work as a draft; enforce grid, palette, clusters, and frames with pixel tools |
| Tileable surface or tileset | Produce a small family; repair seams; test repeated 3×3 before approval |
| UI art or icons | Keep text and interaction code-native; prefer SVG/vector for simple geometric symbols |
| 3D model or material | Use concepts as reference; author/clean in a DCC; validate topology, UVs, scale, pivots, and LODs |
| No generation/edit tool is available | Build the brief and manifest; source licensed assets or keep explicit greybox placeholders |
| Audio asset | Route sound implementation and mixing to audio-design; still track source/license in the manifest |
Image-generation handoff
When a capable image generation or editing tool is installed, use it for live visual creation. If
the workspace exposes an imagegen skill, read and follow it for the actual generation/edit call;
this skill owns the game-art brief, constraints, normalization, and acceptance gates. Generation
APIs usually bill per image: tell the user the batch size and ask before large or repeated runs.
Build prompts from these blocks:
ROLE/PURPOSE: production asset for [gameplay role]
SUBJECT: [specific object/character and action]
VIEW: [orthographic/top-down/side/three-quarter], [camera and facing]
ART DIRECTION: [shape language], [palette roles], [materials], [edge treatment]
GAME-SCALE READ: [silhouette and focal details that must survive at WxH]
TECHNICAL OUTPUT: [dimensions/aspect], [transparent or scene background], [frame/slot count]
LOCKS: preserve [identity, proportions, palette, costume, lighting, line weight]
EXCLUDE: text, labels, mockup frames, scenery, duplicate objects, cropped edges, signatures
For edits, say both what changes and what stays fixed. Ask for transparent output through the tool's native transparency option when available, then verify the alpha channel—prompt wording alone does not guarantee transparency.
Raster QA recipes
Inspect constraints and emit a machine-readable report. --require-alpha also fails an RGBA file
whose every pixel is opaque, which is how a baked background or fake checkerboard shows up:
python scripts/asset_report.py assets/player-idle.png \
--expect-size 64x64 --require-alpha --max-colors 48 --json
Build a nearest-neighbor contact sheet over a checkerboard:
python scripts/build_preview_sheet.py output/player/*.png \
--out output/player-preview.png --columns 4 --cell-size 192
Run script paths relative to this skill directory, or resolve the installed skill path first.
Both scripts need Python and Pillow (see Version notes). Install the only dependency with
python -m pip install -r scripts/requirements.txt when it is not already available.
asset_report.py exits 1 when any explicit constraint fails, so it can gate a build.
Quality gates
- Cohesion: related assets share palette roles, line/edge treatment, view, light direction, scale, and detail density.
- Gameplay read: silhouettes and state changes remain clear at native resolution, in motion, and over real backgrounds.
- Technical fit: exact size/frame count, usable alpha, stable anchors/pivots, correct color space/filtering, no clipped content, no accidental labels or baked mockup chrome.
- Animation: identity, volume, proportions, costume, facing, and baseline do not drift; timing and anticipation read in an in-engine preview.
- Tiles/backgrounds: required edges tile without seams; repetition is tolerable; parallax layers have intentional depth and no baked collision cues.
- 3D: transforms, scale, pivot, normals, UVs, materials, topology, rig, collision proxies, LODs, and runtime format are checked rather than inferred from a render.
- Rights: every shipped file has recorded provenance and terms compatible with the project.
Do not call an asset production-ready from a prompt result alone. Approval requires the relevant technical checks plus an in-engine or native-scale visual inspection.
Gotchas
- Fake transparency. Image models often paint a grey-and-white checkerboard or a flat
backdrop instead of alpha.
asset_report.py --require-alphafails a file with no alpha channel or with every pixel opaque; remove the background with a real matte, never by trusting the prompt. - Halos around cut-outs. Semi-transparent edges keep the old background colour and show as a light or dark fringe in the engine. Bleed edge colour outward into transparent pixels (or export premultiplied when the engine expects it) and check over a contrasting background.
- Pseudo pixel art. Generated "pixel art" has blended in-between colours, off-grid pixels and
hundreds of near-duplicates. Downscale to the true grid with nearest-neighbour, quantize to
the palette, then check
--max-colors. - Mipmaps or linear filtering on pixel art blur and shimmer it. Nearest filtering, no mipmaps, integer camera scale.
- Sprite sheets that do not slice. Generated grids drift by a few pixels per cell and frames change size. Re-slice onto a fixed canvas with one shared anchor (usually bottom-centre) before import.
- Character drift between frames or batches (head size, outline width, palette). Edit from the approved seed each time instead of from the last output.
- Texture bleeding at tile and atlas edges under zoom or mipmaps. Pad or extrude tiles by 1-2 px and test a 3x3 repeat at the game's camera scale.
- Block-compression sizes. BC and ETC2 compress in 4x4 blocks (ASTC in its chosen block size); other dimensions get padded, or rejected by stricter paths such as WebGL compressed textures. Keep textures at multiples of 4, and powers of two where mipmaps are needed.
- Normal and data maps imported as sRGB shade wrongly. Only colour textures are sRGB.
- Generated text and logos come out as near-letters and invite trademark trouble. Keep text code-native and remove marks before approval.
- Licences that do not survive packing. Attribution and terms must follow the asset through atlases and renames; record them in the manifest, not in a filename.
- Undeclared AI content. Steam's Content Survey asks about AI-generated content that ships
with the game; answer it from the manifest's
source.kind, so record generation per asset.
References
- For visual-system decisions and maintaining consistency, read
references/art-direction.md. - For sprites, animation strips, tiles, backgrounds, UI art, and engine import settings, read
references/raster-pipeline.md. - For concept-to-mesh, textures, glTF/GLB, LOD, collision, and runtime validation, read
references/three-d-pipeline.md. - For licenses, generated-media records, and provenance, read
references/provenance.md.
Related skills
game-ui-uxfor layout, navigation, readability, and accessible interaction.game-feel,shader-programming, andaudio-designfor presentation after the source art fits.- Engine import/rendering skills such as
godot-tilemap,unity-tilemap-2d,pixijs-rendering, andthreejs-gltf-loading.
Version notes
Last verified: 2026-11-11.
- Scripts: Python 3.10+ with Pillow 10 to 13 (
scripts/requirements.txtpinsPillow>=10,<14). Pillow 13.0.0 is the latest release and requires Python 3.11+ (https://pypi.org/project/pillow/). - Steam Content Survey: the generative AI section separates pre-generated content (made during development and shipped) from live-generated content (made while the game runs, which also needs a description of guardrails); AI tools used only for efficiency are not its focus (https://partner.steamgames.com/doc/gettingstarted/contentsurvey).
- Engine import settings in
references/raster-pipeline.mdtarget Godot 4.7 and Unity 6.3 LTS.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |