Content
85%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 a disciplined operational spec: four sequenced phases, explicit validation gates ('No citation, no install', corpus-missing STOP), a failure-mode checklist, and copy-paste-ready hook wiring for five hook managers. Its main defect is progressive disclosure — the two referenced branch files are absent from the bundle, so propose-only mode and re-runs cannot actually be executed as shipped — with minor repetition of the no-tag-pre-filter rule keeping conciseness just short of lean.
Suggestions
Ship propose-only.md and re-runs.md inside the skill bundle (e.g., references/) or state their exact location relative to SKILL.md — both are named as required reading for whole modes of operation but do not exist in the bundle, so Phase 3's propose-only branch and the Re-runs branch dead-end.
Consolidate the three statements of the no-tag-pre-filter rule ('never a tag-wide dump', 'There is no tag pre-filter', 'A fired tag earns NOTHING by itself') into one canonical statement plus the Redis example to trim redundant tokens.
State where the corpus's Detect:/Not-when: rule format is documented or include one short example rule, since selection correctness hinges on a file format the skill only describes secondhand.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Quotes: "probe the stack, select ONLY rules whose subject the repo's code actually shows — evidence-cited per rule, never a tag-wide dump — confirm with the user, then install each confirmed rule's corpus prose verbatim" — the body is lean with zero padding and no explanations of concepts Claude already knows. It falls short of anchor 5 because the no-tag-pre-filter rule is stated three times ("never a tag-wide dump", "There is no tag pre-filter", "A fired tag earns NOTHING by itself") and several triple-clause sentences could be tightened; it is above anchor 3 because every repetition carries new operational nuance rather than filler. | 4 / 5 |
Actionability | Quotes: "pre-commit: commands: <rule-name>: run: <gate>", "a `<gate>` line in `.husky/pre-commit`", "`.tsx`/`.jsx` files, JSX or react imports in real source", "`.publish(`, `.send(`, `.create(`…", ".claude/sound/<topic>/<rule-name>.md" — exact paths, copy-paste-ready wiring per hook tool, concrete detection cues, and worked judgment examples (next-auth EmailProvider as email-sending, redis-as-cache vs. store, sqlite as embedded db). Fully executable for an instruction-only skill, so no code-absence penalty applies. | 5 / 5 |
Workflow Clarity | Quotes: "Phase 1 — Probe" through "Phase 4 — Write" give a named, ordered sequence; validation is explicit ("If `corpus/` is missing or unreadable, STOP and say so", "No citation, no install"); error recovery is a feedback loop ("a corrected set replaces yours without argument; redo the affected phases"), and "Failure modes to keep loud" is a five-entry checklist with prescribed responses. This matches anchor 5: clear sequence, explicit validation steps, feedback loops, and a checklist. | 5 / 5 |
Progressive Disclosure | Quotes: "read `propose-only.md` and follow it instead of confirming or writing", "read `re-runs.md` before writing anything" — the deferral strategy is right (conditional branches externalized, one level deep, clearly signaled at point of use), but neither referenced file exists in the bundle (no references/, scripts/, or assets/ directories, no sibling files), so both references dead-end as shipped. Not anchor 4 because broken references are more than a minor organization gap; not anchor 2 because nothing that belongs in separate files is inlined and the signaling is clear. | 3 / 5 |
Total | 17 / 20 Passed |