Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Local evidence retrieval (FlashRAG-style) for VCO/vibe: search protocols/config/skills docs and return citeable snippets with file+line anchors.
.claude/skills/foryourhealth111-pixel-flashrag-evidence/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | -51% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 22% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 21% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 1% | 0% |
Use this skill when you need grounded, citeable evidence from local documentation/configuration to support VCO decisions or recommendations, especially for:
This skill is not a replacement for GitNexus (code dependency graph) or web search. It focuses on local docs and config.
~/.codex/skills/vibe/:protocols/, config/, references/, scripts/router/~/.codex/skills/**/SKILL.md) for tool capability evidencepython C:\Users\羽裳\.codex\skills\flashrag-evidence\scripts\flashrag_evidence.py --query "…" --topk 8bm25s)pwsh C:\Users\羽裳\.codex\skills\vibe\scripts\ruc-nlpir\preflight.ps1install-upstreams.ps1 auto-install path has been removed on purpose.C:\Users\羽裳\.codex\_external\ruc-nlpir\.venv\Scripts\python.exe C:\Users\羽裳\.codex\skills\flashrag-evidence\scripts\flashrag_evidence.py --engine bm25s --query "…" --topk 8--roots to include the project workspace--topkrg -n on the most likely file(s)The script prints ranked evidence items:
path + line (1-based) for quick navigationscore for rankingsnippet (short, safe to quote)| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 5,611 | 4,961 | -12% | 1 | 1 | 0% | 955 | 1,038 | +9% | 0 | 0 | — |
case-02 | fail→fail | 4,311 | 5,598 | +30% | 1 | 1 | 0% | 270 | 984 | +264% | 0 | 0 | — |
case-03 | fail→fail | 6,131 | 4,980 | -19% | 1 | 1 | 0% | 1,009 | 1,016 | +1% | 0 | 0 | — |
case-04 | fail→fail | 14,565 | 8,101 | -44% | 1 | 1 | 0% | 2,409 | 2,128 | -12% | 0 | 0 | — |
case-05 | pass→pass | 10,057 | 8,032 | -20% | 1 | 1 | 0% | 1,753 | 2,151 | +23% | 0 | 0 | — |
case-06 | fail→pass | 15,948 | 3,048 | -81% | 1 | 1 | 0% | 2,521 | 1,244 | -51% | 0 | 0 | — |
case-07 | fail→pass | 18,697 | 2,912 | -84% | 1 | 1 | 0% | 1,002 | 1,221 | +22% | 0 | 0 | — |
case-08 | fail→pass | 11,386 | 2,323 | -80% | 1 | 1 | 0% | 1,854 | 1,130 | -39% | 0 | 0 | — |
case-09 | fail→pass | 12,586 | 2,736 | -78% | 1 | 1 | 0% | 1,008 | 1,223 | +21% | 0 | 0 | — |
case-10 | fail→fail | 9,579 | 1,929 | -80% | 1 | 1 | 0% | 1,473 | 1,010 | -31% | 0 | 0 | — |
case-11 | fail→pass | 8,729 | 4,611 | -47% | 1 | 1 | 0% | 1,578 | 1,589 | +1% | 0 | 0 | — |
case-12 | pass→pass | 12,140 | 5,517 | -55% | 1 | 1 | 0% | 2,112 | 1,717 | -19% | 0 | 0 | — |
case-13 | pass→pass | 9,044 | 2,971 | -67% | 1 | 1 | 0% | 1,629 | 1,214 | -25% | 0 | 0 | — |
case-14 | fail→pass | 9,237 | 3,312 | -64% | 1 | 1 | 0% | 1,483 | 1,364 | -8% | 0 | 0 | — |
case-15 | fail→pass | 12,243 | 2,254 | -82% | 1 | 1 | 0% | 2,148 | 1,073 | -50% | 0 | 0 | — |
case-16 | fail→pass | 12,779 | 4,605 | -64% | 1 | 1 | 0% | 2,407 | 1,736 | -28% | 0 | 0 | — |
case-17 | fail→pass | 10,977 | 1,642 | -85% | 1 | 1 | 0% | 1,904 | 1,011 | -47% | 0 | 0 | — |
case-18 | fail→pass | 4,947 | 2,408 | -51% | 1 | 1 | 0% | 816 | 1,180 | +45% | 0 | 0 | — |
case-19 | pass→pass | 11,261 | 5,289 | -53% | 1 | 1 | 0% | 1,813 | 1,575 | -13% | 0 | 0 | — |
case-20 | pass→pass | 11,969 | 4,909 | -59% | 1 | 1 | 0% | 1,842 | 1,571 | -15% | 0 | 0 | — |
case-21 | fail→pass | 8,068 | 4,640 | -42% | 1 | 1 | 0% | 1,376 | 1,491 | +8% | 0 | 0 | — |
case-22 | fail→pass | 14,130 | 4,395 | -69% | 1 | 1 | 0% | 2,450 | 1,110 | -55% | 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 18 counted toward the lift figure. The other 4 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 +55 percentage points is the difference between those two pass rates over the 18 comparable cases.
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.