Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →기존 references/<id>/DESIGN.md(엄격히 작성된 산문)에서 머신리더블 `tokens:` 블록(DTCG-lite: colors/typography/rounded/spacing/shadow/components)을 역추적(prose-derived)해 frontmatter에 backfill. token↔prose 정합성 게이트로 검증하고 token-status 체크리스트를 갱신. 배치(기본 10개)로 며칠에 나눠 실행. '토큰 백필', 'tokens 블록 채워', 'X에 토큰 추가', '토큰 배치 돌려', '남은 레퍼런스 토큰화' 류에 트리거. 신규 reference의 토큰은 omd-add-reference Phase 4.5가 담당(여긴 기존 ref 전용).
.claude/skills/kwakseongjae-omd-token-backfill/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -64% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -75% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -78% | 0% |
Status: CORE_V2_CATALOG_WRITE_BLOCKED.
Do not add or rewrite YAML tokens: blocks. Historical token data remains a dual-read migration input; new structured tokens belong in the canonical System Graph and must project to a vendor-neutral DESIGN.md Core v2 file.
Resume only after the catalog graph/evidence package, token closure validator, registry and quality readers, and dropped-zero legacy migration are available. Until then this skill performs no writes and suggests no plausible fallback.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 11,356 | 3,837 | -66% | 1 | 1 | 0% | 2,308 | 830 | -64% | 0 | 0 | — |
case-02 | fail→pass | 7,456 | 4,143 | -44% | 1 | 1 | 0% | 1,211 | 971 | -20% | 0 | 0 | — |
case-03 | fail→pass | 14,241 | 3,509 | -75% | 1 | 1 | 0% | 3,036 | 749 | -75% | 0 | 0 | — |
case-04 | fail→pass | 9,084 | 2,788 | -69% | 1 | 1 | 0% | 1,289 | 649 | -50% | 0 | 0 | — |
case-05 | fail→pass | 15,358 | 3,153 | -79% | 1 | 1 | 0% | 3,180 | 690 | -78% | 0 | 0 | — |
case-06 | fail→pass | 10,380 | 2,543 | -76% | 1 | 1 | 0% | 1,828 | 589 | -68% | 0 | 0 | — |
case-07 | fail→pass | 10,475 | 3,121 | -70% | 1 | 1 | 0% | 1,682 | 691 | -59% | 0 | 0 | — |
case-08 | fail→pass | 15,016 | 6,028 | -60% | 1 | 1 | 0% | 2,355 | 784 | -67% | 0 | 0 | — |
case-09 | fail→pass | 8,449 | 2,800 | -67% | 1 | 1 | 0% | 1,339 | 633 | -53% | 0 | 0 | — |
case-10 | fail→pass | 3,830 | 2,609 | -32% | 1 | 1 | 0% | 693 | 611 | -12% | 0 | 0 | — |
case-11 | fail→pass | 28,333 | 7,571 | -73% | 1 | 1 | 0% | 1,348 | 870 | -35% | 0 | 0 | — |
case-12 | fail→pass | 14,060 | 2,498 | -82% | 1 | 1 | 0% | 1,145 | 566 | -51% | 0 | 0 | — |
case-13 | fail→pass | 13,686 | 4,698 | -66% | 1 | 1 | 0% | 2,050 | 960 | -53% | 0 | 0 | — |
case-14 | pass→pass | 9,002 | 2,512 | -72% | 1 | 1 | 0% | 1,342 | 596 | -56% | 0 | 0 | — |
case-15 | pass→pass | 25,160 | 2,827 | -89% | 1 | 1 | 0% | 1,586 | 487 | -69% | 0 | 0 | — |
case-16 | fail→pass | 19,289 | 1,386 | -93% | 1 | 1 | 0% | 890 | 319 | -64% | 0 | 0 | — |
case-17 | pass→pass | 14,698 | 4,937 | -66% | 1 | 1 | 0% | 2,358 | 942 | -60% | 0 | 0 | — |
case-18 | pass→pass | 17,392 | 2,127 | -88% | 1 | 1 | 0% | 2,709 | 434 | -84% | 0 | 0 | — |
case-19 | fail→fail | 7,078 | 7,428 | +5% | 1 | 1 | 0% | 1,420 | 1,593 | +12% | 0 | 0 | — |
case-20 | pass→fail | 8,300 | 3,641 | -56% | 1 | 1 | 0% | 1,619 | 794 | -51% | 0 | 0 | — |
case-21 | pass→pass | 12,613 | 7,800 | -38% | 1 | 1 | 0% | 2,125 | 1,775 | -16% | 0 | 0 | — |
case-22 | pass→pass | 18,878 | 16,319 | -14% | 1 | 1 | 0% | 2,972 | 2,759 | -7% | 0 | 0 | — |
case-23 | fail→pass | 9,952 | 3,893 | -61% | 1 | 1 | 0% | 1,553 | 796 | -49% | 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, and 21 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +61 percentage points is the difference between those two pass rates over the 21 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
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 | 8/21/2026 | +59% |
Other measured skills in the registry, with their headline benchmark lift.