Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Author a long-running multi-file rewrite plan that subsequent patch-edit + diff-review + build-test stages will execute, with explicit ownership boundaries and patch-safety guarantees.
.claude/skills/nexu-io-rewrite-plan/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-20 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 27% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -51% | 0% |
Spec §20.3 / §21.3.2: scenario 4 (design → deliverable production code) requires a contract-shaped rewrite plan before the agent starts editing files. The plan is the audit trail every subsequent stage references; without it the patch path slides into "refactor everything that looks wrong" and the build breaks in unreviewable ways.
code/index.json from code-import.token-map/*.json from token-map.textproject-cwd/ └── plan/ ├── plan.md # human-readable narrative ├── ownership.json # { file: '...', layer: 'leaf' | 'shared' | 'route' | 'shell' } ├── steps.json # ordered { id, files[], rationale, risk: 'low' | 'medium' | 'high' }[] └── meta.json # { generatedAt, atomDigest, tokenMapDigest }
steps.json is the input patch-edit reads one entry per iteration. ownership.json is the source of truth for "which files belong to a single component vs. shared infrastructure"; patch-edit refuses to touch a shell-tier file unless the matching step has risk: 'high' and the user explicitly confirmed.
The atom completes when plan.md is non-empty AND steps.json contains at least one step.
single step. Break into per-component steps.
build-test step at the end.Implemented by the daemon runner in apps/daemon/src/plugins/atoms/rewrite-plan.ts. It classifies ownership and writes plan.md, steps.json, and ownership.json.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-20 | fail→pass | 8,186 | 2,230 | -73% | 1 | 1 | 0% | 1,194 | 824 | -31% | 0 | 0 | — |
case-01 | fail→fail | 24,973 | 35,308 | +41% | 1 | 1 | 0% | 4,037 | 803 | -80% | 0 | 0 | — |
case-02 | fail→fail | 34,509 | 35,162 | +2% | 1 | 1 | 0% | 337 | 787 | +134% | 0 | 0 | — |
case-03 | fail→fail | 27,771 | 36,459 | +31% | 1 | 1 | 0% | 4,842 | 817 | -83% | 0 | 0 | — |
case-04 | fail→fail | 4,800 | 13,504 | +181% | 1 | 1 | 0% | 317 | 2,930 | +824% | 0 | 0 | — |
case-05 | pass→fail | 4,602 | 37,052 | +705% | 1 | 1 | 0% | 618 | 854 | +38% | 0 | 0 | — |
case-06 | pass→fail | 12,701 | 5,091 | -60% | 1 | 1 | 0% | 1,805 | 666 | -63% | 0 | 0 | — |
case-15 | fail→pass | 5,675 | 2,108 | -63% | 1 | 1 | 0% | 835 | 794 | -5% | 0 | 0 | — |
case-07 | fail→pass | 17,313 | 16,941 | -2% | 1 | 1 | 0% | 2,751 | 3,496 | +27% | 0 | 0 | — |
case-08 | fail→pass | 10,290 | 6,709 | -35% | 1 | 1 | 0% | 1,909 | 1,583 | -17% | 0 | 0 | — |
case-09 | fail→pass | 18,914 | 7,118 | -62% | 1 | 1 | 0% | 3,385 | 1,649 | -51% | 0 | 0 | — |
case-10 | fail→pass | 8,677 | 38,685 | +346% | 1 | 1 | 0% | 1,325 | 1,885 | +42% | 0 | 0 | — |
case-11 | pass→pass | 12,349 | 5,518 | -55% | 1 | 1 | 0% | 1,847 | 1,273 | -31% | 0 | 0 | — |
case-12 | pass→pass | 7,419 | 2,729 | -63% | 1 | 1 | 0% | 1,034 | 866 | -16% | 0 | 0 | — |
case-13 | pass→pass | 12,182 | 1,718 | -86% | 1 | 1 | 0% | 1,811 | 709 | -61% | 0 | 0 | — |
case-14 | fail→pass | 22,976 | 2,708 | -88% | 1 | 1 | 0% | 3,383 | 900 | -73% | 0 | 0 | — |
case-16 | fail→pass | 11,576 | 2,248 | -81% | 1 | 1 | 0% | 1,687 | 852 | -49% | 0 | 0 | — |
case-17 | fail→pass | 12,043 | 2,174 | -82% | 1 | 1 | 0% | 1,795 | 838 | -53% | 0 | 0 | — |
case-18 | fail→pass | 12,669 | 4,004 | -68% | 1 | 1 | 0% | 2,314 | 1,162 | -50% | 0 | 0 | — |
case-19 | fail→pass | 11,578 | 5,225 | -55% | 1 | 1 | 0% | 1,803 | 1,374 | -24% | 0 | 0 | — |
case-21 | fail→pass | 7,033 | 1,849 | -74% | 1 | 1 | 0% | 998 | 713 | -29% | 0 | 0 | — |
case-22 | fail→pass | 17,176 | 10,381 | -40% | 1 | 1 | 0% | 2,492 | 2,148 | -14% | 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 17 counted toward the lift figure. The other 5 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 +50 percentage points is the difference between those two pass rates over the 17 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.