Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Systematic debugging skill. Use when encountering a bug, test failure, unexpected behavior, or when asked to "investigate", "debug", "diagnose", or "figure out why". Enforces root-cause-first discipline with structured hypothesis tracking.
.claude/skills/blueberrycongee-investigate/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -25% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 374% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -24% | 0% |
Systematic debugging workflow. Do not guess-and-fix — find the root cause first.
Gather all available evidence before forming any hypothesis.
git log --oneline -20 in the affected area for recent changesDeliverable: a structured symptom summary with affected files, error output, and recent changes.
Form up to 3 ranked hypotheses. For each:
Test hypotheses in rank order. For each:
3-strike rule: If all 3 hypotheses fail, step back and re-collect. Do not keep guessing. Re-read the code path more carefully, expand the search area, or check for environmental factors (config, dependencies, platform).
Only after root cause is confirmed:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-18 | pass→pass | 12,875 | 9,145 | -29% | 1 | 1 | 0% | 1,967 | 1,887 | -4% | 0 | 0 | — |
case-01 | fail→fail | 23,814 | 10,573 | -56% | 1 | 1 | 0% | 192 | 723 | +277% | 0 | 0 | — |
case-02 | fail→fail | 4,280 | 4,865 | +14% | 1 | 1 | 0% | 227 | 693 | +205% | 0 | 0 | — |
case-03 | fail→fail | 4,998 | 3,493 | -30% | 1 | 1 | 0% | 211 | 727 | +245% | 0 | 0 | — |
case-04 | pass→fail | 12,943 | 4,799 | -63% | 1 | 1 | 0% | 2,697 | 685 | -75% | 0 | 0 | — |
case-05 | fail→fail | 1,778 | 2,635 | +48% | 1 | 1 | 0% | 290 | 673 | +132% | 0 | 0 | — |
case-06 | pass→pass | 17,648 | 24,808 | +41% | 1 | 1 | 0% | 3,197 | 4,732 | +48% | 0 | 0 | — |
case-07 | fail→pass | 7,666 | 3,842 | -50% | 1 | 1 | 0% | 1,252 | 941 | -25% | 0 | 0 | — |
case-08 | fail→pass | 4,701 | 19,710 | +319% | 1 | 1 | 0% | 792 | 3,757 | +374% | 0 | 0 | — |
case-09 | fail→fail | 15,473 | 4,613 | -70% | 1 | 1 | 0% | 2,686 | 1,123 | -58% | 0 | 0 | — |
case-10 | fail→pass | 5,948 | 2,729 | -54% | 1 | 1 | 0% | 935 | 956 | +2% | 0 | 0 | — |
case-11 | pass→pass | 13,083 | 3,357 | -74% | 1 | 1 | 0% | 1,812 | 1,008 | -44% | 0 | 0 | — |
case-12 | pass→pass | 9,881 | 2,638 | -73% | 1 | 1 | 0% | 1,546 | 929 | -40% | 0 | 0 | — |
case-13 | fail→pass | 11,391 | 5,871 | -48% | 1 | 1 | 0% | 1,905 | 1,442 | -24% | 0 | 0 | — |
case-14 | pass→pass | 9,454 | 4,325 | -54% | 1 | 1 | 0% | 1,557 | 1,156 | -26% | 0 | 0 | — |
case-15 | fail→pass | 10,589 | 5,234 | -51% | 1 | 1 | 0% | 1,657 | 1,260 | -24% | 0 | 0 | — |
case-16 | pass→pass | 10,398 | 4,339 | -58% | 1 | 1 | 0% | 1,767 | 1,167 | -34% | 0 | 0 | — |
case-17 | fail→fail | 12,914 | 8,080 | -37% | 1 | 1 | 0% | 1,858 | 1,684 | -9% | 0 | 0 | — |
case-19 | fail→fail | 8,983 | 5,813 | -35% | 1 | 1 | 0% | 1,369 | 914 | -33% | 0 | 0 | — |
case-20 | fail→pass | 12,575 | 4,814 | -62% | 1 | 1 | 0% | 2,018 | 1,171 | -42% | 0 | 0 | — |
case-21 | pass→pass | 10,104 | 4,772 | -53% | 1 | 1 | 0% | 1,508 | 1,187 | -21% | 0 | 0 | — |
case-22 | fail→pass | 12,109 | 4,082 | -66% | 1 | 1 | 0% | 1,846 | 1,173 | -36% | 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 19 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 +27 percentage points is the difference between those two pass rates over the 19 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.