Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →スキル、ルール、エージェント定義が実際に遵守されているかを可視化する——3種類のプロンプト厳格度レベルのシナリオを自動生成し、エージェントを実行し、動作シーケンスを分類し、完全なツール呼び出しタイムラインの遵守率をレポートする
.claude/skills/affaan-m-skill-comply/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -32% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -33% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -25% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -55% | 0% |
通过以下方式测量编码代理是否实际遵循技能、规则或代理定义：
claude -p 并通过 stream-json 捕获工具调用轨迹skills/*/SKILL.md）：工作流技能，如搜索优先、TDD 指南rules/common/*.md）：强制性规则，如 testing.md、security.md、git-workflow.mdagents/*.md）：代理是否在预期时被调用（内部工作流验证尚不支持）/skill-comply <path>bash# Full run uv run python -m scripts.run ~/.claude/rules/common/testing.md # Dry run (no cost, spec + scenarios only) uv run python -m scripts.run --dry-run ~/.claude/skills/search-first/SKILL.md # Custom models uv run python -m scripts.run --gen-model haiku --model sonnet <path>
测量技能/规则是否在提示未明确支持时仍被遵循。
报告是自包含的，包括：
对于熟悉钩子的用户，报告还包含针对合规性较低的步骤的钩子提升建议。此为参考信息——主要价值在于合规性本身的可见性。
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 6,730 | 2,329 | -65% | 1 | 1 | 0% | 1,124 | 759 | -32% | 0 | 0 | — |
case-02 | fail→pass | 13,767 | 4,282 | -69% | 1 | 1 | 0% | 2,406 | 1,195 | -50% | 0 | 0 | — |
case-03 | fail→fail | 5,004 | 6,272 | +25% | 1 | 1 | 0% | 338 | 854 | +153% | 0 | 0 | — |
case-04 | fail→fail | 11,228 | 17,607 | +57% | 1 | 1 | 0% | 1,950 | 828 | -58% | 0 | 0 | — |
case-05 | fail→pass | 7,586 | 2,245 | -70% | 1 | 1 | 0% | 1,416 | 947 | -33% | 0 | 0 | — |
case-06 | fail→pass | 6,908 | 2,587 | -63% | 1 | 1 | 0% | 1,264 | 942 | -25% | 0 | 0 | — |
case-07 | fail→pass | 12,866 | 3,275 | -75% | 1 | 1 | 0% | 2,479 | 1,113 | -55% | 0 | 0 | — |
case-08 | fail→pass | 12,061 | 5,904 | -51% | 1 | 1 | 0% | 2,364 | 1,612 | -32% | 0 | 0 | — |
case-09 | fail→pass | 19,122 | 6,065 | -68% | 1 | 1 | 0% | 3,414 | 1,622 | -52% | 0 | 0 | — |
case-10 | fail→fail | 17,913 | 5,225 | -71% | 1 | 1 | 0% | 3,305 | 749 | -77% | 0 | 0 | — |
case-11 | pass→fail | 11,441 | 5,450 | -52% | 1 | 1 | 0% | 1,982 | 745 | -62% | 0 | 0 | — |
case-16 | fail→pass | 8,213 | 2,011 | -76% | 1 | 1 | 0% | 1,362 | 868 | -36% | 0 | 0 | — |
case-12 | fail→pass | 10,888 | 7,085 | -35% | 1 | 1 | 0% | 1,918 | 1,035 | -46% | 0 | 0 | — |
case-13 | pass→pass | 14,249 | 7,043 | -51% | 1 | 1 | 0% | 2,674 | 1,835 | -31% | 0 | 0 | — |
case-14 | pass→pass | 13,062 | 8,358 | -36% | 1 | 1 | 0% | 2,101 | 1,927 | -8% | 0 | 0 | — |
case-15 | fail→pass | 14,004 | 4,513 | -68% | 1 | 1 | 0% | 2,611 | 1,350 | -48% | 0 | 0 | — |
case-17 | pass→pass | 14,393 | 4,441 | -69% | 1 | 1 | 0% | 2,486 | 1,240 | -50% | 0 | 0 | — |
case-18 | fail→pass | 11,228 | 2,514 | -78% | 1 | 1 | 0% | 2,027 | 899 | -56% | 0 | 0 | — |
case-19 | fail→pass | 11,174 | 5,758 | -48% | 1 | 1 | 0% | 1,971 | 1,501 | -24% | 0 | 0 | — |
case-20 | fail→pass | 4,364 | 4,009 | -8% | 1 | 1 | 0% | 596 | 1,159 | +94% | 0 | 0 | — |
case-21 | fail→fail | 9,257 | 11,739 | +27% | 1 | 1 | 0% | 1,888 | 2,937 | +56% | 0 | 0 | — |
case-22 | pass→fail | 14,074 | 5,451 | -61% | 1 | 1 | 0% | 2,623 | 783 | -70% | 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. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 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 +50 percentage points is the difference between those two pass rates over the 18 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.
Other measured skills in the registry, with their headline benchmark lift.