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.
A highly actionable, well-structured review skill: the workflow is fully sequenced with validation feedback loops, and detail is properly externalized into one-level-deep reference files. Its one weakness is density — repeated authorization/verification rules and long multi-clause sentences that could be consolidated and tightened.
Suggestions
Consolidate the authorization rules stated three times across 'Outcome Contract' (Authorization), 'Worktree Safety Preflight', and 'Autofix Routing' into a single stated-once section referenced from the others.
Break the 60+ word multi-clause sentences (e.g. the Hard Rules 'No unverified claims' and 'Dependency changes' entries) into short separate rules so each condition is individually readable.
Merge the duplicated verdict-invalidation rules — 'the old verdict expires' in Get the Diff and 'Any fix made during review invalidates the pre-fix verdict' in Autofix Routing — into one re-freeze rule stated once.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body teaches no concepts Claude already knows, but authorization and verification rules recur across sections ("Outcome Contract" Authorization, "Worktree Safety Preflight", "Autofix Routing", Hard Rules) and many sentences are dense multi-clause runs (e.g. the 60+ word Hard Rules sentences) that could be tightened. This fits 'mostly efficient but could be tightened' better than the minor-trim-only level 4, while clearly above the padded level 2. | 3 / 5 |
Actionability | Concrete, copy-paste commands cover the common path: "git status --short --branch -uall", "gh run view <id> --json status,conclusion", "git worktree add --detach <known-good-commit>", and "bash <skill-base-dir>/scripts/run-tests.sh", plus a fill-in sign-off template. The default review flow is fully executable from the body alone; specialist detail is deliberately one load away. | 5 / 5 |
Workflow Clarity | A clear sequence (preflight → project context → mode pick → Get the Diff → Scope → findings → sign-off) with explicit validation checkpoints and feedback loops: "Any fix made during review invalidates the pre-fix verdict. Re-freeze the baseline, re-run the check", the dirty-checkout isolation procedure, and a ledger-style sign-off checklist. Destructive and ship operations are heavily gated, so no validation cap applies. | 5 / 5 |
Progressive Disclosure | The Mode Picker table gives well-signaled, conditional, one-level-deep references ("load references/mode-triage.md" etc.), and all nine referenced files plus three scripts exist in the bundle. Content is appropriately split (modes, personas, patterns, surfaces externalized; shared core inline) with easy navigation; the only cross-links are between sibling reference files, not deep nesting. | 5 / 5 |
Total | 18 / 20 Passed |