Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Tactic: Read one paper by Keshav's three-pass method — a shallow skim, a contribution-grasping full read, then a deep virtual re-implementation. Use when the goal is understanding a paper rather than extracting a fixed schema.
.claude/skills/yogsoth-ai-keshav-three-pass/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 65% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -2% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-21 | ✗→✓ | ▲ Improved | 4% | 0% |
Read one paper in three passes of increasing depth. The outputs accumulate as prose rather than fixed fields; use a different tactic when cross-paper alignment matters more than understanding.
paper-fetch with paper_ref. Stop on not_found. Createcontext/papers/<dir>/keshav-three-pass/ on success.
first-pass-skim with source_path and meta_path. Write01-first-pass-skim.md, recording read_deeper in frontmatter.
read_deeper is false, stop by default. Continue only on explicitcaller override and record gate_overridden: true.
second-pass-grasp with the paths and skim_notes; write02-second-pass-grasp.md.
third-pass-deep-read with the paths and grasp_summary; write03-third-pass-deep-read.md.
Do not collapse pass 3 into a recap of pass 2. It must surface implicit assumptions, virtual re-implementation mismatches, and concrete improvements.
textcontext/papers/<timestamp>-<title-slug>/ source.md source.meta.json keshav-three-pass/ 01-first-pass-skim.md 02-second-pass-grasp.md 03-third-pass-deep-read.md
Each output carries sop, tactic, and written_at frontmatter. Report the gate outcome, core claim, most consequential implicit assumption, unresolved flags, and all output paths.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 54,048 | 65,779 | +22% | 1 | 1 | 0% | 7,525 | 12,394 | +65% | 0 | 0 | — |
case-02 | fail→fail | 28,418 | 17,305 | -39% | 1 | 1 | 0% | 4,152 | 1,047 | -75% | 0 | 0 | — |
case-03 | fail→fail | 37,438 | 14,752 | -61% | 1 | 1 | 0% | 5,362 | 804 | -85% | 0 | 0 | — |
case-04 | fail→pass | 20,722 | 14,526 | -30% | 1 | 1 | 0% | 2,631 | 2,118 | -19% | 0 | 0 | — |
case-05 | fail→fail | 29,260 | 46,193 | +58% | 1 | 1 | 0% | 4,320 | 8,664 | +101% | 0 | 0 | — |
case-06 | fail→fail | 23,210 | 92,027 | +296% | 1 | 1 | 0% | 2,924 | 8,655 | +196% | 0 | 0 | — |
case-07 | fail→pass | 29,360 | 28,730 | -2% | 1 | 1 | 0% | 4,980 | 4,895 | -2% | 0 | 0 | — |
case-08 | fail→fail | 84,125 | 18,192 | -78% | 1 | 1 | 0% | 12,960 | 881 | -93% | 0 | 0 | — |
case-09 | fail→fail | 20,692 | 37,247 | +80% | 1 | 1 | 0% | 2,646 | 3,220 | +22% | 0 | 0 | — |
case-10 | pass→fail | 17,127 | 20,096 | +17% | 1 | 1 | 0% | 2,139 | 1,293 | -40% | 0 | 0 | — |
case-11 | pass→pass | 53,438 | 14,543 | -73% | 1 | 1 | 0% | 8,251 | 2,063 | -75% | 0 | 0 | — |
case-12 | fail→fail | 53,804 | 16,601 | -69% | 1 | 1 | 0% | 11,064 | 2,526 | -77% | 0 | 0 | — |
case-13 | fail→fail | 34,726 | 50,133 | +44% | 1 | 1 | 0% | 5,194 | 8,638 | +66% | 0 | 0 | — |
case-14 | fail→fail | 17,050 | 62,705 | +268% | 1 | 1 | 0% | 310 | 8,665 | +2695% | 0 | 0 | — |
case-15 | fail→pass | 48,947 | 34,821 | -29% | 1 | 1 | 0% | 7,620 | 6,117 | -20% | 0 | 0 | — |
case-16 | fail→fail | 33,554 | 16,662 | -50% | 1 | 1 | 0% | 4,921 | 911 | -81% | 0 | 0 | — |
case-17 | fail→fail | 16,639 | 20,077 | +21% | 1 | 1 | 0% | 1,782 | 3,179 | +78% | 0 | 0 | — |
case-18 | fail→fail | 20,773 | 22,966 | +11% | 1 | 1 | 0% | 2,996 | 879 | -71% | 0 | 0 | — |
case-19 | fail→fail | 24,110 | 23,778 | -1% | 1 | 1 | 0% | 3,528 | 3,534 | +0% | 0 | 0 | — |
case-20 | fail→fail | 34,239 | 50,967 | +49% | 1 | 1 | 0% | 6,477 | 8,660 | +34% | 0 | 0 | — |
case-21 | fail→pass | 40,383 | 37,358 | -7% | 1 | 1 | 0% | 6,038 | 6,270 | +4% | 0 | 0 | — |
case-22 | pass→fail | 24,251 | 36,761 | +52% | 1 | 1 | 0% | 3,117 | 6,065 | +95% | 0 | 0 | — |
case-23 | fail→fail | 71,409 | 15,923 | -78% | 1 | 1 | 0% | 12,226 | 827 | -93% | 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 15 counted toward the lift figure. The other 8 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 +13 percentage points is the difference between those two pass rates over the 15 comparable cases. 3 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.