Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Investigate application or repository source code with search_file and read_file to explain behavior, trace root causes of runtime or query errors, and answer with precise file citations.
.claude/skills/frankchen021-source-code-inspection/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | 551% | 0% |
| case-05 | ✓→✗ | ▼ Worse | -58% | 0% |
| case-08 | ✓→✗ | ▼ Worse | -65% | 0% |
| case-15 | ✓→✗ | ▼ Worse | 9% | 0% |
| case-17 | ✓→✗ | ▼ Worse | -12% | 0% |
Use this skill when the user wants to:
UNKNOWN_TABLEsearch_file first to locate relevant files and line numbers before reading content.read_file only for targeted sections; do not read entire files or large ranges.no matches found, refine the search and retry before giving up.Always use this exact format for source-file citations:
[[file:path/to/file.ts]][[file:path/to/file.ts#L12]][[file:path/to/file.ts#L12-L34]]Use repo-relative paths and line anchors whenever possible. Do not use normal markdown links for source-code citations. Do not wrap file citation tokens in backticks, inline code, or fenced code blocks in your final answer. Use them as plain citation tokens in normal prose or list items so the UI can render them as clickable links.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-11 | fail→fail | 5,078 | 8,245 | +62% | 1 | 1 | 0% | 770 | 744 | -3% | 0 | 0 | — |
case-01 | fail→fail | 3,013 | 4,700 | +56% | 1 | 1 | 0% | 271 | 596 | +120% | 0 | 0 | — |
case-02 | fail→fail | 18,407 | 5,478 | -70% | 1 | 1 | 0% | 2,600 | 623 | -76% | 0 | 0 | — |
case-03 | fail→fail | 14,561 | 6,885 | -53% | 1 | 1 | 0% | 2,823 | 786 | -72% | 0 | 0 | — |
case-04 | fail→fail | 3,677 | 7,373 | +101% | 1 | 1 | 0% | 640 | 1,547 | +142% | 0 | 0 | — |
case-05 | pass→fail | 12,121 | 7,678 | -37% | 1 | 1 | 0% | 1,982 | 837 | -58% | 0 | 0 | — |
case-06 | fail→fail | 4,942 | 12,213 | +147% | 1 | 1 | 0% | 207 | 1,899 | +817% | 0 | 0 | — |
case-07 | pass→pass | 10,443 | 3,493 | -67% | 1 | 1 | 0% | 1,715 | 902 | -47% | 0 | 0 | — |
case-08 | pass→fail | 14,820 | 6,176 | -58% | 1 | 1 | 0% | 2,252 | 784 | -65% | 0 | 0 | — |
case-09 | pass→pass | 6,442 | 2,757 | -57% | 1 | 1 | 0% | 1,056 | 829 | -21% | 0 | 0 | — |
case-10 | fail→pass | 1,778 | 6,220 | +250% | 1 | 1 | 0% | 217 | 1,412 | +551% | 0 | 0 | — |
case-12 | fail→fail | 28,335 | 8,249 | -71% | 1 | 1 | 0% | 4,890 | 779 | -84% | 0 | 0 | — |
case-13 | fail→fail | 12,172 | 7,037 | -42% | 1 | 1 | 0% | 1,718 | 688 | -60% | 0 | 0 | — |
case-14 | pass→pass | 4,832 | 2,868 | -41% | 1 | 1 | 0% | 756 | 817 | +8% | 0 | 0 | — |
case-15 | pass→fail | 4,480 | 6,758 | +51% | 1 | 1 | 0% | 611 | 669 | +9% | 0 | 0 | — |
case-16 | fail→fail | 2,835 | 7,121 | +151% | 1 | 1 | 0% | 255 | 766 | +200% | 0 | 0 | — |
case-17 | pass→fail | 5,472 | 11,422 | +109% | 1 | 1 | 0% | 800 | 701 | -12% | 0 | 0 | — |
case-18 | pass→pass | 4,549 | 2,547 | -44% | 1 | 1 | 0% | 683 | 726 | +6% | 0 | 0 | — |
case-19 | fail→fail | 4,503 | 5,358 | +19% | 1 | 1 | 0% | 622 | 570 | -8% | 0 | 0 | — |
case-20 | pass→fail | 23,409 | 5,696 | -76% | 1 | 1 | 0% | 4,869 | 581 | -88% | 0 | 0 | — |
case-21 | pass→fail | 15,745 | 6,892 | -56% | 1 | 1 | 0% | 2,847 | 575 | -80% | 0 | 0 | — |
case-22 | pass→fail | 16,334 | 5,808 | -64% | 1 | 1 | 0% | 3,092 | 564 | -82% | 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 6 counted toward the lift figure. The other 16 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 6 comparable cases. 7 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.