Content
92%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is an exemplary lean advisory skill: token-efficient, fully executable, with a clearly sequenced workflow and built-in verification gates. Its single real defect is bundle-level — the referenced lens.md is missing entirely, so the one progressive-disclosure promise the skill makes cannot be fulfilled.
Suggestions
Ship the referenced `lens.md` in the bundle or delete the "Lens-eligible for the `pr-reviewer` agent" line — the file does not exist next to SKILL.md, so the claimed auto-load on deep reviews cannot happen.
If per-axis detection heuristics or additional worked examples are added later, place them one level deep in a clearly signaled reference file rather than growing SKILL.md, keeping the current overview-first shape.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Every section earns its place: the axis table ("a hand-rolled helper duplicating an export; a bespoke pattern next to a shared one"), the drop rules, the three-part materiality bar, and the output template all carry judgment criteria Claude does not already know. There is no concept tutorial, no library primer, no padded prose — the closing line "silence on a sound approach costs nothing; a weak 'better way' costs trust" encodes the core decision rule rather than over-explaining, so this is anchor 5, not anchor 4's "minor instances of over-explanation". | 5 / 5 |
Actionability | Fully executable for a report-only skill: the exact diff command ("git diff <base>...HEAD, base default origin/main"), a copy-paste output template with per-field slots ("**Now:** <the current approach, one line, with file:line>"), the exact terminal string ("print exactly `Approach: optimal — no materially better approach found.`"), and a worked ✓/✗ pair grounded in file:line. This matches anchor 5's "copy-paste ready… specific examples cover the common cases"; the scoring note for instruction-only skills confirms no code absence penalty applies. | 5 / 5 |
Workflow Clarity | Steps 0–3 are numbered and unambiguous (scope → score four axes → drop what another lens owns → materiality bar → output), with explicit verification gates built in: "grep before you claim one exists" and "stated in two sentences with file:line evidence you read, not a guess". The materiality bar ("Material / Grounded / Worth the switch") is a pre-output checklist and the exact optimal-string removes terminal-state ambiguity — matching anchor 5. The destructive/batch validation cap does not apply since the skill is "Report only; never edit files". | 5 / 5 |
Progressive Disclosure | The inline structure is strong (clear headed sections, ~64 well-organized lines, nothing that should be split out), but the skill's only external reference — "Lens-eligible for the `pr-reviewer` agent: [`lens.md`](./lens.md) auto-loads on `deep` reviews" — points to a file that does not exist in the bundle (no lens.md, and no references/, scripts/, or assets/ directories), so the advertised navigation and deep-review integration are broken. Structure alone reads as anchor 5 and the dangling link as a clear defect, landing between anchors; the false reference is more than anchor 4's "minor organization gaps", so 3. | 3 / 5 |
Total | 18 / 20 Passed |