Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →This skill should be used only when the user explicitly asks to use `$ralph-specum-refactor`, or explicitly asks Ralph Specum in Codex to revise spec artifacts after implementation learnings.
.claude/skills/tzachbon-ralph-specum-refactor/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 724% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 64% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -54% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -35% | 0% |
You are a coordinator, not a refactor specialist -- delegate spec revision to a refactor-specialist sub-agent.
.current-specrequirements.md, design.md, tasks.md.progress.md and existing spec files.refactor-specialist sub-agent. Pass .progress.md, existing spec files, and implementation learnings. The sub-agent identifies what changed, what stayed accurate, and what is obsolete. Do NOT revise spec files yourself.[P] tasks, [VERIFY] tasks, VE tasks, and epic constraints when relevant.requirements.mddesign.mdtasks.md.progress.md.approve current artifactrequest changescontinue to implementationcontinue to implementation as approval of the updated spec files.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 7,816 | 6,476 | -17% | 1 | 1 | 0% | 1,098 | 705 | -36% | 0 | 0 | — |
case-02 | fail→pass | 2,106 | 15,592 | +640% | 1 | 1 | 0% | 278 | 2,290 | +724% | 0 | 0 | — |
case-03 | fail→fail | 7,609 | 4,255 | -44% | 1 | 1 | 0% | 1,056 | 676 | -36% | 0 | 0 | — |
case-04 | fail→pass | 5,398 | 6,193 | +15% | 1 | 1 | 0% | 823 | 1,352 | +64% | 0 | 0 | — |
case-05 | pass→pass | 11,202 | 6,618 | -41% | 1 | 1 | 0% | 1,701 | 1,003 | -41% | 0 | 0 | — |
case-06 | pass→pass | 13,891 | 9,862 | -29% | 1 | 1 | 0% | 2,140 | 1,857 | -13% | 0 | 0 | — |
case-07 | pass→pass | 8,587 | 4,601 | -46% | 1 | 1 | 0% | 1,220 | 1,051 | -14% | 0 | 0 | — |
case-08 | fail→pass | 14,627 | 5,568 | -62% | 1 | 1 | 0% | 1,989 | 1,226 | -38% | 0 | 0 | — |
case-09 | fail→pass | 8,801 | 1,624 | -82% | 1 | 1 | 0% | 1,230 | 571 | -54% | 0 | 0 | — |
case-10 | pass→pass | 5,367 | 5,450 | +2% | 1 | 1 | 0% | 874 | 1,284 | +47% | 0 | 0 | — |
case-11 | fail→pass | 10,245 | 3,766 | -63% | 1 | 1 | 0% | 1,406 | 914 | -35% | 0 | 0 | — |
case-12 | pass→pass | 10,238 | 2,705 | -74% | 1 | 1 | 0% | 1,464 | 733 | -50% | 0 | 0 | — |
case-13 | fail→fail | 5,396 | 8,221 | +52% | 1 | 1 | 0% | 320 | 777 | +143% | 0 | 0 | — |
case-14 | fail→pass | 13,740 | 5,580 | -59% | 1 | 1 | 0% | 2,180 | 1,321 | -39% | 0 | 0 | — |
case-15 | pass→pass | 11,710 | 9,390 | -20% | 1 | 1 | 0% | 1,636 | 1,865 | +14% | 0 | 0 | — |
case-16 | pass→fail | 12,807 | 3,887 | -70% | 1 | 1 | 0% | 1,750 | 893 | -49% | 0 | 0 | — |
case-17 | pass→fail | 7,394 | 5,543 | -25% | 1 | 1 | 0% | 1,146 | 1,300 | +13% | 0 | 0 | — |
case-18 | fail→pass | 6,818 | 13,477 | +98% | 1 | 1 | 0% | 1,059 | 1,842 | +74% | 0 | 0 | — |
case-19 | fail→pass | 8,505 | 3,431 | -60% | 1 | 1 | 0% | 1,141 | 900 | -21% | 0 | 0 | — |
case-20 | fail→pass | 16,834 | 14,520 | -14% | 1 | 1 | 0% | 3,263 | 2,764 | -15% | 0 | 0 | — |
case-21 | fail→fail | 27,761 | 9,319 | -66% | 1 | 1 | 0% | 4,627 | 665 | -86% | 0 | 0 | — |
case-22 | fail→fail | 6,023 | 21,244 | +253% | 1 | 1 | 0% | 453 | 2,500 | +452% | 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 +32 percentage points is the difference between those two pass rates over the 18 comparable cases. 2 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.