Content
63%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.
A rigorous, highly actionable policy manual with excellent mode dispatch, verification checklists, and feedback loops, weakened by redundant restatement of the same rules across sections and by references to files (examples/README.md, examples/*.json, detector/validate.js) that are not in the bundle. The progressive-disclosure pattern for the main catalog is sound, but the body is far more than an overview.
Suggestions
State each rule once and cross-reference it: protected-content preservation, pass accounting, and model-only labeling each appear in 3-4 sections; consolidating them into the Editing contract (with the other sections pointing back) would substantially cut the 317-line body without losing operational content.
Fix or remove the broken bundle references: either ship examples/README.md, example style configs, and detector/validate.js, or explicitly mark them as not bundled (as the body already hedges for the detector with 'If the repo ships the detector engine') so the --style config path is not documented as the 'Preferred' option while being unusable.
Compress the 'What this skill is and isn't' research-citation paragraph to its one-line takeaway ('signals, not proof') and move the study citations into references/patterns.md alongside the other sourcing.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The 317-line body avoids explaining known concepts (good), but restates the same rules several times: protected-content preservation appears in the Editing contract ("Quotations, attributed passages, code, tables, URLs, paths, identifiers, frontmatter... retain their content"), again in edit mode ("Follow the editing contract for protected material"), and again in Output format ("Residuals cover the whole supplied text, including protected regions"); pass-accounting rules likewise appear in both "Iterate to convergence" ("A no-op uses none") and Output format ("Do not count returning the unchanged source as an editing pass"), and the model-only labeling rule is repeated across three sections. Not 2 because the density is redundant-but-novel policy rather than padding about things Claude already knows; not 4 because the repetition is substantial enough that the body could be cut roughly in half. | 3 / 5 |
Actionability | Mostly executable guidance: concrete commands with flags and exit codes ("node scripts/normalize-quotes.js <rewritten-prose> --reference <original> --write", "node scripts/check-style.js <file> --config <path>" with "exit 0 clean / 1 hard violation / 2 tool error", "node detector/validate.js --residual-policy warn <original> <rewritten>"), explicit mode triggers, and a four-item Verification checklist. Gaps keep it from 5: detector/validate.js and the examples/ configs referenced in the --style section are not present in the bundle, and the <rewritten-prose> scratch-file convention is under-specified. Not 3 because the main flows (rewrite/detect/edit, marks pass, style check) have real, runnable commands. | 4 / 5 |
Workflow Clarity | Each mode has a numbered step sequence with explicit validation checkpoints ("After editing, re-read the file", "Before delivery, compare the final text with the source") and feedback loops with budgeted repair ("If review finds an unauthorized change, repair it only within the remaining editing budget; otherwise report the unresolved failure"), plus explicit stop reasons and pass accounting — the destructive-capable edit-in-place mode is validated, so no cap applies. It falls short of 5 because the full pipeline must be assembled from six scattered sections (Editing contract, Modes, Invocation, marks pass, Output format, Tone calibration) and the edit-mode mechanical check is only 'optional, recommended' — minor gaps consistent with the anchor-4 midpoint-plus. | 4 / 5 |
Progressive Disclosure | The main reference is well handled — a reference-loading block up top directs reading references/patterns.md, which exists and is one level deep — but the body also points to bundle paths that do not exist: "Schema and rationale: examples/README.md", bare-name resolution against "examples/<name>.json", and "node detector/validate.js" (no examples/ or detector/ directory ships), so navigation fails for the --style and validator features, and the 317-line body inlines heavy policy that is more than an overview. Not 4 because broken references are a concrete navigation failure beyond minor organization gaps; not 2 because the primary reference exists, is clearly signaled, and nothing is buried two levels deep. | 3 / 5 |
Total | 14 / 20 Passed |