Content
70%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 well-structured, token-conscious index with an exemplary gated workflow, explicit validation, and clear checklists. Its two weaknesses are consequential: the bundle ships only references/bloat-patterns.md while the entire execution layer lives in rules/*.md files that are absent, leaving the skill an index to missing content, and there is moderate table-level redundancy.
Suggestions
Ship the rules/*.md files referenced by the Workflow and Required Reading tables (hard-rules, measurement, classification, invocation-review, and the three mode files) in the bundle — without them, Phases 1–3 have no executable detail and the load-on-demand design points at files that don't resolve.
Merge the 'Workflow' and 'Required Reading by Phase' tables into one table with columns for phase, rule file, gate, and load cue — they currently duplicate the same phase→file mapping and account for the largest avoidable repetition in the body.
Inline one-line fallbacks for the critical criteria (e.g. the token estimate formula and the 6-line paragraph threshold already appear; add the hot-path/cold-path definition in one line) so a basic audit run can proceed even when the rule files are unavailable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Efficient and dense — tables for modes and phases, a "thin index" note, and deliberate non-duplication ("References… token-economics.md and progressive-disclosure.md… do not duplicate that guidance here"). Minor trimmable redundancy: the Workflow and Required Reading tables repeat the same phase→rule-file mapping, and the <10k-chars / never-delete rules are stated in three places. Not 3 — nothing is padded or explains concepts Claude already knows. | 4 / 5 |
Actionability | The orchestration layer is concrete (mode-detection table, exact output template "Mode: audit / Target: /abs/path/to/CLAUDE.md (43,012 chars / ~10,750 tokens)", gated phases, checklist), but the actual per-phase procedures are delegated to rules/*.md files that are not present in the bundle — "rules/measurement.md", "rules/classification.md", and the three mode files are all missing, so how to measure, classify, and run each mode is unavailable. Missing key details is more than the 'minor gaps' of level 4. | 3 / 5 |
Workflow Clarity | Clear Phase 0–4 sequence with explicit validation and feedback loops: Phase 4 "Before/after metrics shown; no content lost in trim or extract", a preservation-invariant gate, "If you cannot find a destination, abort and ask", diff approval for trim, and a Definition of Done checklist. The destructive/batch cap does not apply because validation checkpoints are explicit; not 4 — recovery paths and checklists are all present. | 5 / 5 |
Progressive Disclosure | The thin-index, load-on-demand design is well-signaled ("Detailed rules live in rules/*.md and load on demand"), but scored against the actual bundle the navigation chain breaks: 7 of 8 referenced paths (rules/hard-rules.md, rules/measurement.md, rules/classification.md, rules/invocation-review.md, and the three mode files) do not exist in the bundle, and references/bloat-patterns.md itself links to a missing ../rules/classification.md. Not 4 — 'minor organization gaps' does not cover most references being unresolvable; not 2 because structure and signaling within SKILL.md itself are strong and nothing is buried. | 3 / 5 |
Total | 15 / 20 Passed |