Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when implementing a user- or reviewer-prescribed code change (including review comments with suggested fixes or options), especially when the requested edit may be risky, incomplete, ambiguous, or misaligned with TiDB correctness and compatibility constraints.
.claude/skills/pingcap-tidb-change-instruction-critic/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-22 | ✗→✓ | ▲ Improved | -18% | 0% |
| case-07 | ✓→✗ | ▼ Worse | -71% | 0% |
| case-13 | ✓→✓ | = Same ✓ | -18% | 0% |
| case-15 | ✓→✓ | = Same ✓ | -1% | 0% |
| case-16 | ✓→✓ | = Same ✓ | -24% | 0% |
Use this skill before implementing a concrete approach proposed by the user or a reviewer, including review comments with fix instructions or solution options.
Treat the proposed method as a hypothesis to check against the intended outcome and existing contracts. Explicit user requirements remain constraints; do not silently change scope or SQL semantics to make an implementation easier.
Investigate questions that code, tests, or the existing conversation can answer. Risk, a possible semantic impact, or an initially unclear validation plan does not by itself require another confirmation.
Ask a concise question only when investigation leaves an unresolved requirement, a contract tradeoff the user must decide, or work outside the authorized scope. Explain the concrete decision and its consequence. Pause the dependent change while continuing independent authorized work.
When the approach is sound and authorized, proceed without a separate alignment checkpoint. Explain material departures from a suggested method or remaining tradeoffs; a fixed options report before every edit is unnecessary.
Keep the diff focused on the agreed intent. Follow AGENTS.md for regression coverage and validation, and report the resulting behavior, supporting evidence, and residual risks.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 16,764 | 17,438 | +4% | 1 | 1 | 0% | 373 | 721 | +93% | 0 | 0 | — |
case-02 | fail→fail | 21,822 | 9,691 | -56% | 1 | 1 | 0% | 309 | 730 | +136% | 0 | 0 | — |
case-03 | fail→fail | 6,230 | 5,793 | -7% | 1 | 1 | 0% | 303 | 654 | +116% | 0 | 0 | — |
case-04 | fail→fail | 44,030 | 8,266 | -81% | 1 | 1 | 0% | 8,271 | 656 | -92% | 0 | 0 | — |
case-05 | fail→fail | 25,970 | 12,379 | -52% | 1 | 1 | 0% | 5,044 | 663 | -87% | 0 | 0 | — |
case-06 | fail→fail | 3,766 | 13,204 | +251% | 1 | 1 | 0% | 525 | 543 | +3% | 0 | 0 | — |
case-07 | pass→fail | 14,948 | 11,825 | -21% | 1 | 1 | 0% | 2,153 | 629 | -71% | 0 | 0 | — |
case-08 | fail→fail | 12,197 | 7,883 | -35% | 1 | 1 | 0% | 1,843 | 1,539 | -16% | 0 | 0 | — |
case-09 | fail→fail | 11,005 | 7,719 | -30% | 1 | 1 | 0% | 1,585 | 796 | -50% | 0 | 0 | — |
case-10 | fail→fail | 12,068 | 8,878 | -26% | 1 | 1 | 0% | 1,756 | 1,624 | -8% | 0 | 0 | — |
case-11 | fail→fail | 11,850 | 12,335 | +4% | 1 | 1 | 0% | 1,783 | 1,535 | -14% | 0 | 0 | — |
case-12 | fail→fail | 6,996 | 12,570 | +80% | 1 | 1 | 0% | 1,056 | 1,041 | -1% | 0 | 0 | — |
case-13 | pass→pass | 7,923 | 5,544 | -30% | 1 | 1 | 0% | 1,168 | 959 | -18% | 0 | 0 | — |
case-14 | fail→fail | 15,795 | 7,954 | -50% | 1 | 1 | 0% | 2,391 | 808 | -66% | 0 | 0 | — |
case-15 | pass→pass | 15,391 | 16,941 | +10% | 1 | 1 | 0% | 2,205 | 2,177 | -1% | 0 | 0 | — |
case-16 | pass→pass | 12,739 | 6,953 | -45% | 1 | 1 | 0% | 1,959 | 1,488 | -24% | 0 | 0 | — |
case-17 | pass→pass | 6,518 | 9,177 | +41% | 1 | 1 | 0% | 910 | 1,165 | +28% | 0 | 0 | — |
case-18 | fail→fail | 11,799 | 16,164 | +37% | 1 | 1 | 0% | 1,822 | 2,324 | +28% | 0 | 0 | — |
case-19 | fail→fail | 24,189 | 12,551 | -48% | 1 | 1 | 0% | 2,180 | 2,177 | -0% | 0 | 0 | — |
case-20 | fail→fail | 8,944 | 10,995 | +23% | 1 | 1 | 0% | 1,621 | 1,293 | -20% | 0 | 0 | — |
case-21 | fail→fail | 7,300 | 4,913 | -33% | 1 | 1 | 0% | 1,044 | 942 | -10% | 0 | 0 | — |
case-22 | fail→pass | 14,815 | 9,874 | -33% | 1 | 1 | 0% | 2,070 | 1,690 | -18% | 0 | 0 | — |
case-23 | pass→pass | 24,987 | 15,731 | -37% | 1 | 1 | 0% | 1,909 | 1,999 | +5% | 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 14 counted toward the lift figure. The other 9 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 0 percentage points is the difference between those two pass rates over the 14 comparable cases. 4 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.