Content
81%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 unusually rigorous instruction document: three clearly triggered modes, exact commands with exit codes, an explicit editing contract, pass-budget accounting, and a demanding verification protocol with feedback loops. Its weaknesses are redundancy — pass-budget and verification rules are restated across several sections, and the research framing runs long — and a handful of referenced paths (examples/, detector/) that do not exist in this bundle. Strong overall, with tightening and path fixes as the remaining improvements.
Suggestions
Consolidate the editing-pass budget and verification rules into a single authoritative section; they are currently restated in 'Iterate to convergence', the 'Rewrite mode' output-format paragraphs, and the Verification item list, which a reader must reconcile.
Compress the false-positive research framing (Liang et al., Jabarian & Imas citations) and meta-explanations like 'Why it belongs here rather than in the pattern catalog' to one or two operational sentences each; the operative guidance ('signals, not proof; don't use for consequential decisions') survives without the full literature review.
Resolve the dangling bundle references: either ship the examples/ config directory (including examples/README.md) and detector/validate.js referenced by the --style and mechanical-check sections, or explicitly mark those paths as available only in the full repository so every path named in SKILL.md resolves.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly purposeful policy with no explaining of concepts Claude already knows, but includes unnecessary repetition and could be tightened: the editing-pass budget rules are restated in "Iterate to convergence," again in "Rewrite mode" ("repair it only within the remaining editing budget"), and again across the Verification paragraphs; the "Never inject" list self-identifies duplication ("Covered as a detection rule under Speculative scenario openers; listed again here"); and the ~200-word research-citation framing (Liang et al., Jabarian & Imas) plus meta-justifications like "Why it belongs here rather than in the pattern catalog" could each compress to a sentence or two. Fits the score-3 anchor (mostly efficient, some unnecessary explanation) rather than 2 because the density is redundancy, not fluffy padding, and every repeated passage still carries some operational signal. | 3 / 5 |
Actionability | Fully executable, copy-paste-ready commands with documented exit semantics: `node scripts/normalize-quotes.js <rewritten-prose> --reference <original> --write`, `node scripts/check-style.js <file> --config <path>` (exit 0 clean / 1 hard violation / 2 tool error), and `node detector/validate.js --residual-policy warn <original> <rewritten>`; mode triggers are given as quoted user phrases; the output contract names exact sections (Final rewrite, Changes, Verification with Editing passes / Checks / Residuals / Stop reason); and the pattern catalog lists concrete offending phrases per tier. Matches the score-5 anchor for an instruction-only skill with concrete, specific guidance covering the common cases. | 5 / 5 |
Workflow Clarity | Each mode has a numbered sequence (rewrite: Audit → Rewrite → Summarize; edit: Read → Edit in place → Verify) with explicit validation checkpoints and feedback loops: "After editing, re-read the file and report whether another justified in-scope edit remains," the before-delivery source comparison with repair "only within the remaining editing budget," the no-op decision gate, and mandatory residual/stop-reason reporting with model-only labeling when checks cannot run. Matches the score-5 anchor (clear sequence, explicit validation, error-recovery loops); the in-place edit mode is not missing validation, so no cap applies. | 5 / 5 |
Progressive Disclosure | Good structure overall: the 870-line pattern catalog is properly externalized to references/patterns.md behind a clearly signaled one-level-deep loading block ("read [references/patterns.md] in full"), and referenced scripts (normalize-quotes.js, check-style.js) exist and are invoked with exact commands. Held at 4 rather than 5 because several referenced paths do not resolve in the actual bundle — "examples/<name>.json", "examples/README.md" (schema and rationale), and "node detector/validate.js" have no examples/ or detector/ directory here — a minor organization gap against the 'clear overview, easy navigation' anchor. Not 3: the main catalog split is appropriate and well-signaled, and the missing detector path is hedged ("If the repo ships the detector engine"). | 4 / 5 |
Total | 17 / 20 Passed |