Content
66%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 well-engineered workflow: clear phases, explicit validation checkpoints, a fully specified output template, and careful handling of ambiguous states (zero-hit scans, unset config, person-only facts). Its main weakness is token efficiency — design-rationale blockquotes and a cert_tier rule stated three times pad roughly a third of the content without adding executable instruction.
Suggestions
Cut or compress the design-rationale blockquotes (the 'wrong test' aside, the 'template wins in practice' note, and the long zero-hit preamble) into single-line rules — the operative instruction survives without the argument for it.
State the cert_tier gating rule once (at the top) and have Phase 4 reference it in one line instead of re-explaining the four values and re-arguing the gating a second and third time.
Trim the zero-hit ambiguity section to its operative rule (establish and report a denominator per scan; report NOT ASSESSED with a reason) — the 20-line justification can be reduced to the two-line legend already defined.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body interleaves multiple design-rationale blockquotes that argue for the skill's own design rather than instruct ("Emitting every item for every platform trains the reader to skip the list", "is the wrong test: it treats an itch project and a steam project identically", "When a rule and a template disagree, the template wins in practice"), and the cert_tier rule is stated three times (top-of-skill rule, Phase 4 gate, closing blockquote). This is noticeably verbose with padded sections. It is not 1 because nothing explains concepts Claude already knows and the operational core is dense with real instruction; it is not 3 because roughly a third of the body is justification of the rules rather than the rules themselves. | 2 / 5 |
Actionability | Guidance is concrete and executable throughout: exact scan terms ("TODO, FIXME, HACK", "placeholder, temp_, WIP_"), an exact output line format ("scanned [N] files, [M] hits"), a full copy-paste checklist template, exact AskUserQuestion option labels, and an exact status precedence rule ("First match wins"). Minor gaps hold it at 4 rather than 5: the scans name search strings but no literal commands, and key inputs defer to external docs (effects-map.md, code-root-resolution.md) without inline fallbacks. | 4 / 5 |
Workflow Clarity | Phases 1-6 are clearly sequenced with explicit validation checkpoints: the denominator-before-every-scan rule, the [?] vs [ ] distinction carried into the output, NOT ASSESSED handling for unresolved config, first-match-wins status resolution with full precedence, a dedicated person-only confirmation phase (4b) with a defined answer taxonomy, and a permission gate before any file creation ("Create nothing before the answer"). This is a checklist-driven complex process with feedback loops, matching the top anchor. | 5 / 5 |
Progressive Disclosure | No bundle files exist; external doc references (.claude/docs/automation-modes.md, effects-map.md, code-root-resolution.md, config-resolution.md) are one level deep, clearly signaled, and gated with unusually good hygiene ("read only that section (Grep its heading, then a bounded Read); the file as a whole is not a runtime input"). Phase headings make navigation easy. It falls short of 5 because the ~235-line checklist template and the interleaved meta-commentary live inline in a single monolithic file rather than an overview-plus-references split. | 4 / 5 |
Total | 15 / 20 Passed |