Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Build, audit, and revise physically plausible Blender optical-table figures from 2D schematics and manufacturer CAD. Use for Thorlabs-compatible mirrors, lens mounts, beamsplitters, posts, fiber launches, apertures, detectors, optical topology, fail-closed evidence, or Nature-style photonics renders.
.claude/skills/k-telux-thorlabs-blender-optical-path/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -7% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 78% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 13% | 0% |
Turn a measurement requirement or 2D optical schematic into a physically explainable, independently auditable Blender scene.
BLOCKED or UNVERIFIED when real geometry or evidence is missing. File existence, imported CAD, process success, labels, AABB contact, and self-reported text are not proof.First distinguish a new measurement design from a reconstruction, an existing-scene correction, or a presentation-only rerender. Record which inputs are design authority, reusable assets, or visual references. When the user asks for a new path at the quality of an old example, derive a new topology and build from an empty scene; the old scene is not the starting model. For optics-only requests, exclude electrical/data/circuit visualization while retaining optical detectors and their mechanical supports.
scripts/workflow_ledger.py; record every gate, artifact hash, blocker, invalidation, and next authorized action there.schematic node -> experimental role -> real asset -> optical/fiber/electrical ports -> support path.BLOCKED result rather than an exception or silently collapsed entry.active_states or a hashed deterministic expansion, and every declared state has explicit ray templates.Use low-cost diagnostic previews to review device detail, beam readability, and framing. Label them as previews with physical gaps explicit. After physical gates pass, produce final editorial views with restrained metals, readable black-anodized edges, modest glass, thin emissive beam cores, soft halos, and uncluttered labels. Preserve geometry and material-slot semantics through presentation changes. Inspect component close-ups and every declared branch; whole-image sharpness, edge density, or colored-pixel totals alone do not prove fidelity or continuity. Verify the actual requested output dimensions; a 2K preview does not satisfy a 4K delivery.
Once the requested checks pass, freeze one delivery and publish once. Recheck affected dependencies after a change, identity drift, or failed test; do not restart unchanged successful gates or open additional variants for optional polish. Skill documentation releases may record limitations without granting model-release credit; see multi-run-qualification.md.
PASS: every applicable active gate has fresh evidence.PARTIAL/SCOPED: only a declared subset was audited; enumerate blockers and keep final/release approval false.UNVERIFIED: required evidence is missing or cannot distinguish the claim.BLOCKED: a known requirement fails.Never promote PARTIAL/SCOPED to whole-system PASS.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 44,832 | 46,399 | +3% | 1 | 1 | 0% | 8,282 | 9,864 | +19% | 0 | 0 | — |
case-02 | fail→fail | 45,509 | 46,299 | +2% | 1 | 1 | 0% | 8,269 | 8,380 | +1% | 0 | 0 | — |
case-03 | fail→fail | 47,560 | 52,080 | +10% | 1 | 1 | 0% | 8,273 | 9,855 | +19% | 0 | 0 | — |
case-04 | pass→pass | 37,846 | 23,236 | -39% | 1 | 1 | 0% | 7,038 | 5,848 | -17% | 0 | 0 | — |
case-05 | pass→pass | 25,113 | 37,745 | +50% | 1 | 1 | 0% | 3,589 | 6,857 | +91% | 0 | 0 | — |
case-06 | pass→pass | 23,991 | 39,295 | +64% | 1 | 1 | 0% | 3,883 | 8,843 | +128% | 0 | 0 | — |
case-07 | fail→pass | 29,214 | 16,893 | -42% | 1 | 1 | 0% | 4,416 | 4,087 | -7% | 0 | 0 | — |
case-08 | fail→pass | 19,804 | 15,232 | -23% | 1 | 1 | 0% | 2,818 | 3,655 | +30% | 0 | 0 | — |
case-09 | pass→pass | 19,922 | 16,771 | -16% | 1 | 1 | 0% | 2,928 | 4,247 | +45% | 0 | 0 | — |
case-10 | pass→pass | 22,632 | 14,010 | -38% | 1 | 1 | 0% | 3,211 | 3,375 | +5% | 0 | 0 | — |
case-11 | fail→pass | 22,110 | 22,322 | +1% | 1 | 1 | 0% | 2,852 | 5,088 | +78% | 0 | 0 | — |
case-12 | pass→pass | 20,346 | 19,187 | -6% | 1 | 1 | 0% | 3,232 | 4,341 | +34% | 0 | 0 | — |
case-13 | pass→pass | 16,236 | 7,837 | -52% | 1 | 1 | 0% | 2,075 | 2,900 | +40% | 0 | 0 | — |
case-14 | fail→pass | 20,067 | 15,332 | -24% | 1 | 1 | 0% | 2,804 | 3,604 | +29% | 0 | 0 | — |
case-15 | pass→pass | 12,859 | 12,407 | -4% | 1 | 1 | 0% | 2,021 | 3,294 | +63% | 0 | 0 | — |
case-16 | pass→pass | 16,268 | 14,137 | -13% | 1 | 1 | 0% | 2,365 | 3,739 | +58% | 0 | 0 | — |
case-17 | pass→pass | 13,028 | 9,535 | -27% | 1 | 1 | 0% | 1,487 | 3,016 | +103% | 0 | 0 | — |
case-18 | fail→pass | 19,575 | 13,699 | -30% | 1 | 1 | 0% | 2,885 | 3,266 | +13% | 0 | 0 | — |
case-19 | fail→fail | 14,049 | 9,208 | -34% | 1 | 1 | 0% | 2,051 | 2,829 | +38% | 0 | 0 | — |
case-20 | fail→pass | 16,020 | 8,378 | -48% | 1 | 1 | 0% | 2,256 | 2,826 | +25% | 0 | 0 | — |
case-21 | pass→pass | 41,498 | 20,607 | -50% | 1 | 1 | 0% | 2,619 | 4,598 | +76% | 0 | 0 | — |
case-22 | pass→pass | 15,810 | 10,651 | -33% | 1 | 1 | 0% | 1,931 | 3,067 | +59% | 0 | 0 | — |
case-23 | fail→pass | 15,731 | 12,978 | -18% | 1 | 1 | 0% | 2,264 | 3,275 | +45% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted. The headline lift of +30 percentage points is the difference between those two pass rates over the 23 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
| Model | Method | Date | Lift |
|---|---|---|---|
| gemini-3.6-flash | verified | 8/27/2026 | +30% |
| gemini-3.6-flash | verified | 8/22/2026 | +38% |
Other measured skills in the registry, with their headline benchmark lift.