Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →In large or many-file projects, prefer cxs over raw rg/grep to reduce context waste.
.claude/skills/hashgraph-online-rg-budget-search/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-09 | ✗→✓ | ▲ Improved | -46% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 71% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -45% | 0% |
| case-16 | ✗→✓ | ▲ Improved | -67% | 0% |
In large or many-file projects, use MCP cxs(op,args) or CLI cxs before raw rg/grep.
Default flow: find(files_only:true) to identify candidate files, find(paths=[...]) to locate bounded line hits, then use Codex-native reads for the exact file or span.
Use --path or paths for concrete files/directories. Use --scope or scopes only for presets: docs, src, tests, config, analysis, runs, vendor, all.
For small projects, known files, or narrow checks, normal file reads/direct shell tools are fine. If output is truncated, refine query/scope/path or continue files-only pagination with next_page.offset.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-09 | fail→pass | 17,005 | 11,626 | -32% | 1 | 1 | 0% | 2,414 | 1,306 | -46% | 0 | 0 | — |
case-03 | fail→fail | 20,465 | 18,360 | -10% | 1 | 1 | 0% | 623 | 532 | -15% | 0 | 0 | — |
case-01 | fail→fail | 19,607 | 17,390 | -11% | 1 | 1 | 0% | 2,059 | 580 | -72% | 0 | 0 | — |
case-02 | fail→fail | 21,084 | 16,900 | -20% | 1 | 1 | 0% | 2,658 | 575 | -78% | 0 | 0 | — |
case-04 | fail→fail | 14,048 | 18,146 | +29% | 1 | 1 | 0% | 1,281 | 580 | -55% | 0 | 0 | — |
case-05 | pass→fail | 19,810 | 24,747 | +25% | 1 | 1 | 0% | 842 | 670 | -20% | 0 | 0 | — |
case-06 | fail→pass | 25,502 | 14,293 | -44% | 1 | 1 | 0% | 1,041 | 1,785 | +71% | 0 | 0 | — |
case-07 | fail→pass | 15,353 | 25,857 | +68% | 1 | 1 | 0% | 1,775 | 2,138 | +20% | 0 | 0 | — |
case-08 | pass→fail | 6,551 | 17,629 | +169% | 1 | 1 | 0% | 911 | 608 | -33% | 0 | 0 | — |
case-10 | fail→pass | 14,789 | 4,803 | -68% | 1 | 1 | 0% | 1,670 | 919 | -45% | 0 | 0 | — |
case-11 | pass→pass | 18,798 | 12,152 | -35% | 1 | 1 | 0% | 2,086 | 1,343 | -36% | 0 | 0 | — |
case-12 | fail→fail | 7,465 | 18,345 | +146% | 1 | 1 | 0% | 1,134 | 359 | -68% | 0 | 0 | — |
case-13 | fail→fail | 16,141 | 10,368 | -36% | 1 | 1 | 0% | 279 | 406 | +46% | 0 | 0 | — |
case-14 | fail→fail | 17,132 | 10,043 | -41% | 1 | 1 | 0% | 483 | 435 | -10% | 0 | 0 | — |
case-15 | pass→fail | 14,505 | 18,672 | +29% | 1 | 1 | 0% | 1,828 | 640 | -65% | 0 | 0 | — |
case-16 | fail→pass | 13,523 | 6,746 | -50% | 1 | 1 | 0% | 1,323 | 436 | -67% | 0 | 0 | — |
case-17 | fail→fail | 10,291 | 11,421 | +11% | 1 | 1 | 0% | 1,407 | 593 | -58% | 0 | 0 | — |
case-18 | fail→pass | 13,404 | 8,700 | -35% | 1 | 1 | 0% | 2,211 | 617 | -72% | 0 | 0 | — |
case-19 | pass→pass | 11,858 | 3,533 | -70% | 1 | 1 | 0% | 1,799 | 442 | -75% | 0 | 0 | — |
case-20 | pass→pass | 9,719 | 8,975 | -8% | 1 | 1 | 0% | 789 | 738 | -6% | 0 | 0 | — |
case-21 | fail→pass | 6,782 | 2,692 | -60% | 1 | 1 | 0% | 1,224 | 618 | -50% | 0 | 0 | — |
case-22 | fail→fail | 9,828 | 1,455 | -85% | 1 | 1 | 0% | 680 | 409 | -40% | 0 | 0 | — |
case-23 | fail→pass | 4,865 | 4,305 | -12% | 1 | 1 | 0% | 816 | 1,029 | +26% | 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 11 counted toward the lift figure. The other 12 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 +22 percentage points is the difference between those two pass rates over the 11 comparable cases. 6 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.