Content
46%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is an extraordinarily detailed, accurate-feeling contract of renderer invariants with excellent concrete identifiers, but it is a monolithic 330-line wall of prose with almost no sectioning and no progressive disclosure — everything lives inline in SKILL.md. Splitting into topic reference files (landscape, weather, clouds, ghost boards, shadows/lighting) with a concise core contract would fix both the conciseness and structure problems at once.
Suggestions
Split the body into one-level-deep reference files (e.g. references/landscape.md, references/weather-particles.md, references/clouds.md, references/ghost-boards.md) and keep SKILL.md to the core renderer contract (~50 lines) with clearly signaled pointers.
Add section headers for the major rule groups (render targets, geometry caches, batching, culling, shadows/lighting, weather) so the rule catalog is navigable without reading all 330 lines.
Expand the Validation section into a feedback loop: state what to do when a check fails (e.g. 'if postTarget references are found, remove them and re-run the syntax check before reload').
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Every bullet is genuinely project-specific knowledge Claude could not infer, so there is no filler about known concepts — but the body is ~330 lines / 43KB of dense run-on bullets (e.g. the multi-paragraph LandscapeEngine and planet-underlay passages) with repeated justifications that should be tightened or moved out of SKILL.md, matching 'noticeably verbose'. Not a 3 because the volume far exceeds what a SKILL.md overview should carry, and not a 1 because nothing is generic padding. | 2 / 5 |
Actionability | Guidance is highly concrete and executable for this codebase: named functions (`optimizeVoxelObjectGroup(...)`, `mergeStaticBaseMeshesByMaterial(...)`, `renderSceneIfReady()`), exact constants (`SUN_OFFSET = (7, 12, 5)`, `WATERFALL_FROTH_SPEED` 0.30), file paths (`01-render-core.js`, `23-particles-clouds.js`), settings keys (`tinyworld:render:*`), and diagnostic commands (`?demo=island-stress&islands=50&stats=1`). It is not a 5 because most rules state invariants and prohibitions rather than copy-paste-ready steps, and a few fixes are described only abstractly ('solve with material/shader treatment or camera/post tuning'). | 4 / 5 |
Workflow Clarity | The body is an unordered rule catalog rather than a sequenced process; the only workflow-like element is the closing Validation list ('Run the inline script syntax check… Confirm renderer.getPixelRatio() is at or below the cap… Confirm no console errors after reload'), which provides checkpoints but no error-recovery loop. This matches 'steps listed but validation gaps / checkpoints missing or implicit' — not a 4 because there is no clear sequence for the main body of rules, and not a 2 because the Validation section does give concrete, ordered checks. | 3 / 5 |
Progressive Disclosure | There is no bundle (no references/, scripts/, or assets/) and nearly no internal structure: ~300 unbroken bullets under a single implicit section, with only two headers ('## Validation', '## Under-occlusion top-content fade + cloud wipe') appearing at the end. Content that clearly belongs in separate reference files (LandscapeEngine rules, weather/particle rules, cloud rules, ghost-board architecture) is entirely inlined, matching 'minimal structure; content that clearly belongs in separate files is inlined'. Not a 1 because the bullets are thematically related and the two headers plus the external pointer to `.codex/skills/tinyworld-mesh-terrain` give minimal navigation. | 2 / 5 |
Total | 11 / 20 Passed |