A game-art director's workflow for ts23d — style guides / art bibles as code (palettes with roles, proportion and scale rules, material and wear conventions, LOD and budget policy), exemplar assets, consistent review with the critique rubric, keeping a whole set of assets (a city, a fleet, a bestiary) visually coherent. Use when defining or enforcing a style across several assets, reviewing a batch of recipes, planning an asset set, or when assets look like they come from different games.
Installs into .claude/skills of the current project.
Sponsors
SSkills DirectorySkills Directory
Know which skills are safe — weekly.
SSkills DirectorySkills Directory
Is that skill safe to install?
Are you the author of Art Director?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/developerz-ai-art-director)
---
name: art-director
description: A game-art director's workflow for ts23d — style guides / art bibles as code (palettes with roles, proportion and scale rules, material and wear conventions, LOD and budget policy), exemplar assets, consistent review with the critique rubric, keeping a whole set of assets (a city, a fleet, a bestiary) visually coherent. Use when defining or enforcing a style across several assets, reviewing a batch of recipes, planning an asset set, or when assets look like they come from different games.
---
# Game art director
Use with [16-art-as-text](../../../docs/idea/16-art-as-text.md), the
[critique section of author-recipe](../author-recipe/SKILL.md#critique-score--8-on-every-criterion-before-done),
[06-qa critique](../../../docs/idea/06-qa.md), [materials](../../../docs/architecture/materials.md)
and the other discipline skills (load the right one per asset). The guide is law until amended;
exceptions are written back into it. Styles are original: never "in the style of" a named game.
## When
More than one asset must look like one world; a set is planned (fleet, street, bestiary); a batch
of recipes needs review; critique scores disagree between assets.
## Workflow
| # | Art-direction stage | ts23d | Exit |
|---|---|---|---|
| 1 | Pillars: 3 words for the look (e.g. "weathered, industrial, warm") and the realism level | style module `recipes/_lib/style/<world>.ts` (code, not a PDF) | pillars + realism level written |
| 2 | Palette with roles | `definePalette({ primary, secondary, accent, metal, ground… })`; existing `palettes.*` for reference | every colour has a role and a share |
| 3 | Proportion and scale rules | constants: heads-per-height, door height, grid `G`, wheel/door scale; `real.measures` per asset | one module every recipe imports |
| 4 | Material and wear conventions | allowed strips, `paint` slot names, `finishes`, `weathered({ age, wet, wear })` presets, `grime` strength | one wear story per world |
| 5 | Budgets and LOD policy per class | shared `budget` constants; `ctx.pick` patterns; texel range for `assertTexelDensity` | budgets table in the style module |
| 6 | Exemplars | 1–3 hero assets built to the guide first; others measured against them | exemplars score ≥ 8 in critique |
| 7 | Production | discipline skills per asset; `defineFamily` for variants; `ctx.use` for sets | every asset imports the style module |
| 8 | Review cadence | `bun run shots`, `t3d critique` per asset, contact sheets side by side; gallery (`bun run site:dev`) | scores ≥ 8, consistent reasons |
| 9 | Enforce | the set in `recipes/sets.ts` with its `defineStyle` bible; `coherence(style, compiled)` (LOD1) flags robust-z outliers; tests per asset for budgets and silhouette | `checkSet(set)` returns [] and a stray asset fails `T3D_STYLE_OUTLIER` |
## Vocabulary → measurable
| Word | Meaning here | Check |
|---|---|---|
| Art bible / style guide | the rules every asset follows, as importable code + a short doc | style module imported by recipes |
| Palette role | primary / secondary / accent / neutral, with approximate area share (e.g. 60 / 30 / 10) | `styleMetrics(c, style).palette`: area share within CIEDE2000 `paletteDeltaE` (lightness forgiven ×2, so shade ramps belong); ≥ 75 % (`MIN_PALETTE_SHARE`) |
| Realism level | real-scale PBR vs stylised exaggeration, declared | dossier + material ranges (`assertValueRange`) |
| Consistency | same texel density, bevel scale, wear strength, palette, value key across a set | `coherence(style, assets)`: per metric robust z (0.6745·(x − median)/MAD, MAD floored) > `outlierZ` (3.5) → `T3D_STYLE_OUTLIER` naming metric and value |
| Style bible | palette roles, texel density, bevel range, wear range, value key as data | `defineStyle({ name, palette, texelDensity?, bevel?, wear?, valueKey?, outlierZ? })` |
| Scene budget | what a scene costs from street, aerial and overview cameras | `sceneBudgetFindings(stats, placements, budget)`: unique meshes, instances, draws (instanced share), on-screen triangles (LOD by projected size), texture bytes → `T3D_SCENE_BUDGET` |
| Family | animals that share one skeleton | `familyMember(WOLF, { family, mass, length, height, … })`: fox (canid), brown bear (ursid) |
| Scale reference | doors, stairs, a 1.8 m human | `t3d blueprint` figure |
| Readability | silhouette first, value before colour, focal contrast | `assertLodSilhouette`; focal check **planned** |
| Exemplar | the asset others are judged against | critique side by side |
| Critique score | 0–10 per criterion: silhouette, proportions, materials, detail, wear, lod, technical; pass ≥ 8 | `critique.json`, recipes gate (warn mode today) |
| Calibrated score | the same 0–10 predicted from a measured metric vector, fitted on CC0 game assets + degraded copies; a self-score may sit at most 1.5 above it, a blind agent-panel median 2.5 (the panel decides the flagship floor); ratchet in `recipes/critique-baseline.json`; flagship floor 8.5 | `t3d critique` prints it; `calibratedScore`; recipes gate (`T3D_CRITIQUE_*`) |
## Quality bar
- [ ] Pillars, palette roles, scale rules, wear and budgets live in one imported module.
- [ ] The set is declared in `recipes/sets.ts` and `checkSet` returns no `T3D_STYLE_OUTLIER`.
- [ ] Every asset in the set passes the same tests (texel range, budgets, silhouette).
- [ ] Contact sheets side by side look like one world: same light, value range, wear level.
- [ ] Exemplars exist and score ≥ 8 on every criterion.
- [ ] Exceptions are recorded in the style module with a reason.
- [ ] Reviews are harsh and specific: each low score names a recipe change.
## Common mistakes
- A style guide as prose only: nothing enforces it, assets drift.
- One reference game as the whole style: it becomes a clone. Use many references, invent the look.
- Mixing realism levels (photoreal crates beside cartoon characters).
- Each asset inventing its own bevel width, wear strength and palette.
- Scoring generously; "QA PASS" treated as "art approved".
## Run order
```sh
bun run t3d build recipes/<name>/<name>.ts --out out # QA per LOD, exit 1 on a finding
bun run t3d shot recipes/<name>/<name>.ts --out renders # READ sheet.png
bun run t3d hero recipes/<name>/<name>.ts --out renders # READ hero.png
bun run t3d blueprint recipes/<name>/<name>.ts --out renders # READ blueprint.png
bun run site:dev # gallery: the set side by side
bun test recipes/_lib/sets.test.ts # every set in recipes/sets.ts coherent; an off-palette asset fails
bun run t3d critique recipes/<name>/<name>.ts --out tmp/critique # READ every PNG, fill critique.json
# ↳ prints `calibrated X / 10 · claimed Y`: score what you see; a claim > 1.5 over X is T3D_CRITIQUE_INFLATED
bun run golden brief recipes/<name>/<name>.ts # blind panel: 3 independent agents each write review-<k>.json
bun run golden attest recipes/<name>/<name>.ts # median → critique.json (reviewer agent-panel; decides the flagship floor)
bun run golden baseline <name> # after bun run shots: ratchet the calibrated baseline up
bun run shots <name> # refresh committed out/
```
## Pairs with
- [concept-artist](../concept-artist/SKILL.md): exploring the look.
- [material-artist](../material-artist/SKILL.md): palette and wear presets.
- [technical-artist](../technical-artist/SKILL.md): budgets per class.
- [presentation-artist](../presentation-artist/SKILL.md): review images.
- Every discipline skill: load the one that matches each asset.
## Tool status
Shipped: `definePalette`/`palettes`, one trim sheet, `defineFamily`, `ctx.use`, shared parts,
`assertTexelDensity`, `assertValueRange`, `assertLodSilhouette`, `t3d critique` + `critique.json`,
gallery site, `defineStyle` + set coherence (`coherence`, `styleMetrics`, `recipes/sets.ts`: `rts`,
`bestiary`), scene budgets (`sceneBudgetFindings`), skeleton families (`familyMember`). Planned /
gaps: the recipes gate's `sets` pass (wiring), coherence on the beauty render's value key (today
vertex colour), focal/value checks, optional CI fresh-eyes reviewer, enforced critique mode (calibrated inflation and
regression are enforced; the flagship floor reports until slice 17), `variations(n)`.
Neutral positioning: the free code-based option, no claims against other tools.