Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Make simple, focused, low-risk code changes with explicit assumptions and measurable verification. Use for implementation, bug fixes, refactoring, configuration changes, or any task that edits source code. Do not use for read-only explanation or research.
.claude/skills/qierkang-karpathy-guidelines/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-16 | ✗→✓ | ▲ Improved | -41% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -61% | 0% |
| case-23 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-19 | ✓→✗ | ▼ Worse | -50% | 0% |
| case-21 | ✓→✗ | ▼ Worse | -80% | 0% |
Use this skill whenever code or runtime configuration will change. Prefer the simplest implementation that is easy to inspect, test, and remove.
AGENTS.md, the owning package README, and only the files needed for the task.package.json, Docker files, and project scripts first.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-05 | fail→fail | 3,873 | 5,626 | +45% | 1 | 1 | 0% | 639 | 545 | -15% | 0 | 0 | — |
case-01 | fail→fail | 4,412 | 10,310 | +134% | 1 | 1 | 0% | 191 | 518 | +171% | 0 | 0 | — |
case-02 | fail→fail | 3,099 | 4,730 | +53% | 1 | 1 | 0% | 157 | 493 | +214% | 0 | 0 | — |
case-03 | fail→fail | 2,466 | 4,373 | +77% | 1 | 1 | 0% | 136 | 487 | +258% | 0 | 0 | — |
case-04 | pass→pass | 16,685 | 12,123 | -27% | 1 | 1 | 0% | 2,766 | 2,401 | -13% | 0 | 0 | — |
case-06 | pass→pass | 36,239 | 10,623 | -71% | 1 | 1 | 0% | 5,575 | 1,436 | -74% | 0 | 0 | — |
case-07 | pass→pass | 8,368 | 1,565 | -81% | 1 | 1 | 0% | 1,412 | 569 | -60% | 0 | 0 | — |
case-08 | pass→pass | 11,454 | 5,172 | -55% | 1 | 1 | 0% | 1,877 | 1,090 | -42% | 0 | 0 | — |
case-09 | pass→pass | 6,908 | 2,176 | -69% | 1 | 1 | 0% | 1,193 | 625 | -48% | 0 | 0 | — |
case-10 | fail→fail | 12,665 | 7,768 | -39% | 1 | 1 | 0% | 1,960 | 1,245 | -36% | 0 | 0 | — |
case-16 | fail→pass | 13,298 | 6,243 | -53% | 1 | 1 | 0% | 2,340 | 1,385 | -41% | 0 | 0 | — |
case-11 | fail→pass | 12,179 | 3,136 | -74% | 1 | 1 | 0% | 1,900 | 734 | -61% | 0 | 0 | — |
case-12 | pass→pass | 10,038 | 2,080 | -79% | 1 | 1 | 0% | 1,524 | 532 | -65% | 0 | 0 | — |
case-13 | pass→pass | 7,164 | 2,583 | -64% | 1 | 1 | 0% | 1,216 | 618 | -49% | 0 | 0 | — |
case-14 | pass→pass | 11,717 | 3,348 | -71% | 1 | 1 | 0% | 1,695 | 740 | -56% | 0 | 0 | — |
case-15 | pass→pass | 11,545 | 4,212 | -64% | 1 | 1 | 0% | 1,777 | 928 | -48% | 0 | 0 | — |
case-17 | pass→pass | 8,062 | 3,100 | -62% | 1 | 1 | 0% | 1,238 | 742 | -40% | 0 | 0 | — |
case-18 | pass→pass | 9,255 | 3,593 | -61% | 1 | 1 | 0% | 1,453 | 841 | -42% | 0 | 0 | — |
case-19 | pass→fail | 9,410 | 3,102 | -67% | 1 | 1 | 0% | 1,524 | 758 | -50% | 0 | 0 | — |
case-20 | pass→pass | 14,170 | 5,306 | -63% | 1 | 1 | 0% | 2,151 | 1,057 | -51% | 0 | 0 | — |
case-21 | pass→fail | 14,106 | 4,806 | -66% | 1 | 1 | 0% | 2,291 | 461 | -80% | 0 | 0 | — |
case-22 | pass→pass | 7,222 | 3,456 | -52% | 1 | 1 | 0% | 1,278 | 826 | -35% | 0 | 0 | — |
case-23 | fail→pass | 6,091 | 3,162 | -48% | 1 | 1 | 0% | 843 | 748 | -11% | 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 18 counted toward the lift figure. The other 5 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 +4 percentage points is the difference between those two pass rates over the 18 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.