Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Keshav's third pass — the heaviest of the three, a full sentence-by-sentence re-read including proofs/derivations, attempting a virtual re-implementation of the paper to surface implicit assumptions and concrete improvement points. Use this after second-pass-grasp, as the terminal step of the Keshav three-pass method, whenever genuine mastery of a paper (not just a summary) is needed. This is not a skippable recap — treat "nothing new to add" as suspicious, not a default outcome.
.claude/skills/yogsoth-ai-third-pass-deep-read/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | 947% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 462% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -6% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-21 | ✓→✓ | ＝ Same ✓ | 27% | 0% |
Keshav's third pass: sentence-by-sentence re-read with proofs/derivations included, attempting virtual re-implementation. The heaviest pass of the three — terminal step of the Keshav cascade.
Subagent — spawned via spawn-agent skill.
third-pass-verify (v1's name)v1's version of this SOP (staged/wechat-article-v1/skills/third-pass-verify/) treated this as a "targeted re-check of uncertain_fields, no-op if none flagged" step — which, per the coverage audit's S2 finding, effectively deleted Keshav's real third pass (a 4-5+ hour re-implementation attempt) and replaced it with a cheap verification step serving v1's own pipeline. This v2 SOP restores the actual third pass; the rename to third-pass-deep-read marks that this is not the same behavior as the old third-pass-verify, even though both sit in the same cascade position.
A genuine re-implementation attempt needs a context that can hold the full paper and reason through design alternatives without being anchored to how pass 2 already framed the contribution.
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 10,384 | 11,293 | +9% | 1 | 1 | 0% | 839 | 1,324 | +58% | 0 | 0 | — |
case-02 | fail→fail | 14,191 | 11,045 | -22% | 1 | 1 | 0% | 1,403 | 1,293 | -8% | 0 | 0 | — |
case-08 | fail→fail | 26,765 | 41,908 | +57% | 1 | 1 | 0% | 3,529 | 6,864 | +95% | 0 | 0 | — |
case-03 | fail→fail | 8,408 | 15,911 | +89% | 1 | 1 | 0% | 524 | 580 | +11% | 0 | 0 | — |
case-04 | fail→fail | 15,058 | 12,062 | -20% | 1 | 1 | 0% | 1,603 | 1,554 | -3% | 0 | 0 | — |
case-05 | fail→fail | 30,344 | 50,316 | +66% | 1 | 1 | 0% | 3,895 | 8,217 | +111% | 0 | 0 | — |
case-06 | fail→fail | 20,943 | 44,137 | +111% | 1 | 1 | 0% | 3,040 | 7,749 | +155% | 0 | 0 | — |
case-07 | fail→fail | 14,445 | 52,934 | +266% | 1 | 1 | 0% | 1,390 | 8,575 | +517% | 0 | 0 | — |
case-09 | fail→fail | 39,212 | 53,418 | +36% | 1 | 1 | 0% | 6,811 | 10,784 | +58% | 0 | 0 | — |
case-10 | fail→pass | 8,149 | 31,861 | +291% | 1 | 1 | 0% | 456 | 4,773 | +947% | 0 | 0 | — |
case-11 | fail→fail | 42,512 | 44,193 | +4% | 1 | 1 | 0% | 7,286 | 8,571 | +18% | 0 | 0 | — |
case-12 | fail→fail | 47,540 | 48,283 | +2% | 1 | 1 | 0% | 8,224 | 8,500 | +3% | 0 | 0 | — |
case-13 | fail→pass | 13,921 | 42,925 | +208% | 1 | 1 | 0% | 1,237 | 6,957 | +462% | 0 | 0 | — |
case-14 | fail→fail | 55,441 | 65,789 | +19% | 1 | 1 | 0% | 10,141 | 8,567 | -16% | 0 | 0 | — |
case-15 | fail→pass | 15,105 | 12,116 | -20% | 1 | 1 | 0% | 1,598 | 1,495 | -6% | 0 | 0 | — |
case-16 | fail→fail | 39,934 | 52,059 | +30% | 1 | 1 | 0% | 5,226 | 8,569 | +64% | 0 | 0 | — |
case-17 | fail→fail | 42,973 | 43,100 | +0% | 1 | 1 | 0% | 7,428 | 8,566 | +15% | 0 | 0 | — |
case-18 | fail→pass | 16,892 | 15,928 | -6% | 1 | 1 | 0% | 1,744 | 2,205 | +26% | 0 | 0 | — |
case-19 | fail→fail | 46,346 | 16,267 | -65% | 1 | 1 | 0% | 8,216 | 615 | -93% | 0 | 0 | — |
case-20 | fail→fail | 31,762 | 84,551 | +166% | 1 | 1 | 0% | 4,722 | 8,571 | +82% | 0 | 0 | — |
case-21 | pass→pass | 18,726 | 18,955 | +1% | 1 | 1 | 0% | 2,192 | 2,779 | +27% | 0 | 0 | — |
case-22 | pass→pass | 24,666 | 27,860 | +13% | 1 | 1 | 0% | 3,179 | 4,269 | +34% | 0 | 0 | — |
case-23 | pass→pass | 15,337 | 18,684 | +22% | 1 | 1 | 0% | 1,768 | 2,575 | +46% | 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 21 counted toward the lift figure. The other 2 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 +17 percentage points is the difference between those two pass rates over the 21 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.