Content
78%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 lean, well-structured review checklist that earns its tokens with repo-specific guardrails (protected surfaces, eval-canary requirements, confirmation boundaries) rather than generic explanation. The main gaps are a few what-not-how verification bullets and unstated error-recovery loops within the checklist steps.
Suggestions
For the vaguer checklist bullets, add the concrete check or command: e.g. how to inspect routing rules via RAG_PROVIDER_MODE, or which script/test verifies schema conformity and safety filters.
State what to do when a check fails (e.g. flag the finding in the PR review comment vs. block the change), giving each checklist step an explicit feedback loop.
Trim the opening sentence, which duplicates the frontmatter description's trigger information.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is an info-dense checklist with repo-specific paths, env vars, and file pointers rather than explanations of concepts Claude already knows. Minor trimming is possible (the opening sentence restates the description, and the protocol prose in section 0 could be tightened), matching anchor 4 rather than the every-token-earns-its-place anchor 5. | 4 / 5 |
Actionability | Gives concrete, executable guidance: exact files ('AGENTS.md # RAG ranking protection', 'docs/rag-behaviour/', 'ragRankingPatterns in scripts/pr-policy.mjs'), env vars ('RAG_PROVIDER_MODE', 'OPENAI_ANSWER_TIMEOUT_MS'), and concrete artifacts to produce ('Write a PR RAG impact: line', 'post one Review record comment'). Some checklist bullets (e.g. hybrid search, safety filters) state what to verify but not how, keeping it below anchor 5. | 4 / 5 |
Workflow Clarity | A clear numbered checklist sequence with real checkpoints: read-before-editing gates, 'Flag the touch to the user before editing', a required live eval-canary before/after pair ('Offline-green alone is not enough'), and the provider-confirmation boundary. It falls short of anchor 5 because recovery/error-handling loops are not spelled out for each check. | 4 / 5 |
Progressive Disclosure | The skill is under 50 lines, has no bundle files, and is organized into well-labeled sections (protocol, protected surfaces, three numbered review areas) — per the rubric's simple-skill exception, well-organized sections alone merit a 5 here. External file references (AGENTS.md, docs/) are repo artifacts, clearly signaled and one level deep. | 5 / 5 |
Total | 17 / 20 Passed |