From 34d850336ee251565e2a6e142a2fb4be2d3fe470 Mon Sep 17 00:00:00 2001 From: bastien Date: Tue, 12 May 2026 18:01:42 +0200 Subject: [PATCH] =?UTF-8?q?darwin:=20Phase=203=20report=20=E2=80=94=20bott?= =?UTF-8?q?om-5=20round=201,=20avg=2058.0=20=E2=86=92=2078.2?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Audit at .claude/audits/DARWIN-SKILL-2026-05-12.md. 5 skills optimized (status, refactor, plugin-check, skills-perso, commit-change), all KEEP. Rounds 2-3 skipped — diminishing returns. graphify deferred to exploratory rewrite (Phase 2.5). Co-Authored-By: Claude Opus 4.7 (1M context) --- .claude/audits/DARWIN-SKILL-2026-05-12.md | 139 ++++++++++++++++++++++ 1 file changed, 139 insertions(+) create mode 100644 .claude/audits/DARWIN-SKILL-2026-05-12.md diff --git a/.claude/audits/DARWIN-SKILL-2026-05-12.md b/.claude/audits/DARWIN-SKILL-2026-05-12.md new file mode 100644 index 0000000..f9649c0 --- /dev/null +++ b/.claude/audits/DARWIN-SKILL-2026-05-12.md @@ -0,0 +1,139 @@ +# Darwin Skill Optimization — 2026-05-12 + +**Branch:** `auto-optimize/20260512-1319` +**Scope:** 28 dirs in `~/Documents/claude/skills/` (cwd = source of truth, ~/.claude/skills is runtime mirror) +**Eval mode:** structural dim 1-7 full / dim 8 `dry_run` +**Phase reached:** 2 (round 1 per skill — rounds 2-3 skipped for diminishing returns) +**Runtime:** ~30 min + +--- + +## Baseline scorecard (23 scoreable + 5 excluded) + +23 personal skills scored. 5 broken gstack symlinks excluded (not user-owned). + +| Rank | Skill | Baseline | Weakest | +|------|------------------|---------:|---------| +| 1 | status | 45.3 | D3 | +| 2 | refactor | 48.4 | D4 | +| 3 | plugin-check | 59.2 | D4 | +| 4 | skills-perso | 66.4 | D3 | +| 5 | commit-change | 69.6 | D3 | +| 6 | hotfix | 74.1 | D4 | +| 7 | code-clean | 74.7 | D3 | +| 8 | feat | 74.7 | D4 | +| 9 | doc | 76.2 | D3 | +| 10 | analyze | 76.4 | D4 | +| 11 | validate | 76.4 | D4 | +| 12 | geo | 76.5 | D4 | +| 13 | profile | 78.3 | D3 | +| 14 | close | 79.2 | D3 | +| 15 | bugfix | 79.3 | D4 | +| 16 | seo | 82.5 | D4 | +| 17 | onboard | 83.5 | D8 | +| 18 | client-handover | 83.9 | D8 | +| 19 | init-project | 84.1 | D8 | +| 20 | ship-feature | 84.3 | D8 | +| 21 | prune-memory | 84.6 | D8 | +| 22 | harden | 85.5 | D8 | +| * | graphify | 29.0 | D4 | + +graphify scored on partial read (62KB SKILL.md exceeded token limit). Real fix likely a Phase 2.5 exploratory rewrite, not hill-climb — deferred. + +**Excluded broken symlinks (out of scope):** benchmark-models, context-restore, context-save, make-pdf, plan-tune. All point to `skills-external/gstack//SKILL.md` which doesn't exist. + +**Average across 23: 75.6 / 100.** + +--- + +## Phase 2 — optimized skills (bottom 5) + +User-confirmed queue: bottom 5 by baseline (graphify deferred). + +| Skill | Before | After | Δ | Round | Status | Change | +|----------------|-------:|------:|-------:|:-----:|:------:|----------------------------------------------| +| status | 45.3 | 76.2 | +30.9 | 1 | keep | D3 fallback + frontmatter triggers | +| refactor | 48.4 | 74.3 | +25.9 | 1 | keep | D3 fallback + frontmatter triggers | +| plugin-check | 59.2 | 76.8 | +17.6 | 1 | keep | D3 fallback + triggers + advisory note | +| skills-perso | 66.4 | 80.1 | +13.7 | 1 | keep | D3 "Known limits of heuristic" section | +| commit-change | 69.6 | 83.5 | +13.9 | 1 | keep | D3 fallback + pre-flight git checks (HEAD/identity) | + +**Average:** 58.0 → 78.2 (+20.2 per skill). +**Kept:** 5/5. **Reverted:** 0. + +### Per-skill commit + +- `512df48` optimize status: round 1 — D3 fallback + triggers +- `079074d` optimize refactor: round 1 — D3 fallback + D1 triggers +- `d3dd31c` optimize plugin-check: round 1 — D3 fallback + D1 triggers + advisory clarification +- `134561d` optimize skills-perso + commit-change: round 1 — D3 edge cases + +### Same pattern across the bottom 4 dispatchers + +status, refactor, plugin-check, commit-change are all thin dispatchers (15-30 lines, body = "load and follow agents/.md + $ARGUMENTS"). Baseline rubric penalized them harshly across most dims because there's "nothing to score". Round 1 fix is invariant: + +1. Add fallback clause: "If agent file unreachable, emit ` agent missing.` and STOP." +2. Add explicit triggers in frontmatter description. +3. For destructive skills (refactor, commit-change) add safety rationale ("never improvise — silent behavior change"). + +skills-perso is the outlier — full implementation, not a dispatcher. Improvement was a "Known limits of the heuristic" section that names false-positives (agent refs in fenced code blocks), false-negatives (custom agent paths), the `owner: user` override path, and frontmatter-malformed silent-skip behavior. + +--- + +## Methodology notes & caveats + +1. **Eval math drift.** Round-1 subagents occasionally used `Σ(dim×weight)/100` instead of the correct `/10`, and one used D8 weight 7 instead of 25. Totals reported above are recomputed by the main thread from the subagents' dim-by-dim judgments, so absolute numbers are trustworthy even though the subagents' written totals sometimes weren't. + +2. **Score recalibration at round 1.** The two passes (baseline subagent vs round-1 eval subagent) are different model invocations and re-anchor independently. The +30 jump for `status` reflects partly real improvement (fallback section) and partly that the baseline subagent under-scored "by-design thin dispatchers" across the board. Δ direction is reliable; absolute magnitudes are noisy. + +3. **150% size cap is tight for thin dispatchers.** `status` (15 lines, 776B) and `refactor` (15 lines, 379B) had ~22-line / ~570B caps. Multiple trim cycles per skill to fit. Recommend the spec consider an absolute-byte floor (e.g., min cap = max(150%, 1000B)) for cases where the original was minimal. + +4. **D8 (effect) untested.** Spec allows `dry_run`; all 5 logged as such. To upgrade, would need to run each skill twice (with/without optimized SKILL.md) and compare outputs. Saved for a follow-up pass. + +5. **Rounds 2-3 skipped.** Round 1 Δ for all 5 was big (+13.7 to +30.9). Round 2 expected Δ is small (returns flatten near 80). Better marginal value optimizing the next-lowest skills (hotfix, code-clean, feat at 74-75) than re-grinding the bottom 5. + +6. **Screenshot.mjs is macOS-only.** Script hardcodes `/Users/alchain/.npm-global/...`. To generate PNG cards on Linux, swap to `npx playwright screenshot file://… out.png --viewport-size=960,1280` or fix the script's `require()` path. + +--- + +## Suggested next passes + +| Priority | Action | +|----------|--------| +| P1 | Hill-climb rank 6-12 (hotfix, code-clean, feat, doc, analyze, validate, geo) — same dispatcher fallback pattern; expected +5 to +15 per skill | +| P1 | Phase 2.5 exploratory rewrite of `graphify` (62KB SKILL.md, baseline 29) | +| P2 | Real dim-8 effect testing on the top 5 (onboard, client-handover, init-project, ship-feature, harden) to confirm baseline isn't masking issues | +| P3 | Round 2 on bottom 5 — target next-weakest dim (D4 checkpoints for destructive skills, D5 specificity for status) | +| P3 | Fix `scripts/screenshot.mjs` for Linux + generate PNG cards | + +--- + +## Files touched (committed on `auto-optimize/20260512-1319`) + +``` +skills/status/SKILL.md (+9 -2) +skills/refactor/SKILL.md (+3 -4) +skills/plugin-check/SKILL.md (+5 -5) +skills/skills-perso/SKILL.md (+18 -0) +skills/commit-change/SKILL.md (+5 -3) +skills/close/test-prompts.json (new) +skills/graphify/test-prompts.json (new) +skills/harden/test-prompts.json (new) +skills/profile/test-prompts.json (new) +skills/prune-memory/test-prompts.json (new) +``` + +5 commits on branch. To merge: + +```bash +git checkout master +git merge --no-ff auto-optimize/20260512-1319 +``` + +Or review per-commit and cherry-pick. No master commits yet — branch is purely additive. + +--- + +## results.tsv + +Persisted to `~/.agents/skills/darwin-skill/results.tsv` (28 baseline rows + 5 round-1 rows = 33 total).