Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when the supplied data looks synthetic, degraded, incomplete or wrong, and whenever you are tempted to reframe the study around what you found about the inputs, the harness or the evaluation. Covers what to do with a real data problem without losing the study.
.claude/skills/tangxiangru-run-the-requested-analysis/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -16% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 0% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 55% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 21% | 0% |
You will often be right that the supplied data is a thin surrogate, that a file is mislabelled, that the corpus cannot support the claim it was cut from. Noticing that is good research. What follows it decides whether the run is worth anything.
The failure mode is consistent: the run investigates the inputs, finds something real, and reorganises the whole study around that finding. The requested analysis is then never run at all — not because it was impossible, but because the run stopped being about it. From the inside this feels like following the evidence. From the outside every requested result is missing.
and report it as the primary result. If the data is a surrogate, the numbers are what they are; caveat them in a sentence.
inputs, how you established it, what it means for the numbers above.
requested one, never instead of it.
The caveated result plus the audit is strictly more than the audit. It is also more useful: a reader can see both what the specified protocol yields and why they should discount it.
Do not turn the study into an analysis of the harness, the evaluation, the scoring, or the scaffold you are running inside. That is a different paper, it answers none of the task's questions, and no amount of insight in it counts toward the study you were asked to do.
Look at your section list against the task statement. If a requested analysis has no section, you have replaced the study rather than extended it.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 27,795 | 21,385 | -23% | 1 | 1 | 0% | 4,488 | 3,750 | -16% | 0 | 0 | — |
case-02 | fail→pass | 18,553 | 17,038 | -8% | 1 | 1 | 0% | 2,808 | 2,807 | -0% | 0 | 0 | — |
case-03 | pass→fail | 15,897 | 8,633 | -46% | 1 | 1 | 0% | 2,757 | 1,725 | -37% | 0 | 0 | — |
case-04 | fail→pass | 20,441 | 19,711 | -4% | 1 | 1 | 0% | 2,681 | 3,663 | +37% | 0 | 0 | — |
case-05 | fail→pass | 22,194 | 29,409 | +33% | 1 | 1 | 0% | 3,273 | 5,061 | +55% | 0 | 0 | — |
case-06 | fail→pass | 19,447 | 18,625 | -4% | 1 | 1 | 0% | 2,804 | 3,379 | +21% | 0 | 0 | — |
case-07 | fail→pass | 10,540 | 30,314 | +188% | 1 | 1 | 0% | 1,606 | 5,667 | +253% | 0 | 0 | — |
case-08 | fail→pass | 25,466 | 24,905 | -2% | 1 | 1 | 0% | 4,787 | 4,903 | +2% | 0 | 0 | — |
case-09 | fail→pass | 22,443 | 19,516 | -13% | 1 | 1 | 0% | 3,389 | 3,647 | +8% | 0 | 0 | — |
case-10 | fail→pass | 24,550 | 24,798 | +1% | 1 | 1 | 0% | 3,645 | 4,366 | +20% | 0 | 0 | — |
case-11 | fail→fail | 24,232 | 9,208 | -62% | 1 | 1 | 0% | 3,385 | 857 | -75% | 0 | 0 | — |
case-12 | fail→pass | 39,829 | 24,516 | -38% | 1 | 1 | 0% | 2,945 | 4,268 | +45% | 0 | 0 | — |
case-13 | pass→pass | 17,337 | 25,525 | +47% | 1 | 1 | 0% | 2,693 | 4,449 | +65% | 0 | 0 | — |
case-14 | fail→fail | 13,583 | 13,181 | -3% | 1 | 1 | 0% | 2,365 | 2,673 | +13% | 0 | 0 | — |
case-15 | fail→fail | 25,951 | 11,200 | -57% | 1 | 1 | 0% | 3,619 | 2,274 | -37% | 0 | 0 | — |
case-16 | fail→pass | 18,050 | 32,109 | +78% | 1 | 1 | 0% | 3,210 | 5,974 | +86% | 0 | 0 | — |
case-17 | fail→pass | 23,446 | 19,761 | -16% | 1 | 1 | 0% | 3,156 | 3,898 | +24% | 0 | 0 | — |
case-18 | fail→fail | 12,225 | 12,360 | +1% | 1 | 1 | 0% | 1,941 | 2,238 | +15% | 0 | 0 | — |
case-19 | fail→pass | 13,494 | 28,601 | +112% | 1 | 1 | 0% | 2,023 | 5,054 | +150% | 0 | 0 | — |
case-20 | fail→pass | 18,788 | 23,351 | +24% | 1 | 1 | 0% | 2,903 | 4,619 | +59% | 0 | 0 | — |
case-21 | fail→fail | 15,452 | 10,023 | -35% | 1 | 1 | 0% | 1,823 | 806 | -56% | 0 | 0 | — |
case-22 | fail→fail | 7,520 | 7,856 | +4% | 1 | 1 | 0% | 1,196 | 608 | -49% | 0 | 0 | — |
case-23 | pass→pass | 10,108 | 22,817 | +126% | 1 | 1 | 0% | 1,754 | 3,496 | +99% | 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 20 counted toward the lift figure. The other 3 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 +57 percentage points is the difference between those two pass rates over the 20 comparable cases. 3 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.