Content
92%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-engineered instruction skill: the workflow is unambiguous with hard validation gates before any outward-facing action, guidance is fully concrete, and detail is cleanly delegated to one-level-deep references. The only cost is mild rhetorical repetition around the gate rules, which is deliberate safety emphasis rather than padding.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense behavioral instruction with no explanations of concepts Claude already knows, but the step 3 gate rules are restated three times ('A plan that exists only in the subagent's reply or in thinking does not exist for the gate', 'The step 3 gate governs every round, not just the first') and a few justificatory flourishes could be trimmed. | 4 / 5 |
Actionability | Concrete commands ('gh auth status', 'git remote get-url origin'), named platform operations (preflight, checkout-pr, reply-to-thread), an exact subagent dispatch spec, a concrete plan path ('review/pr-<N>-comments-plan.md'), and a worked commit example cover even edge cases (empty working set, dispatch unavailable, pre-commit failure). | 5 / 5 |
Workflow Clarity | Six numbered steps with a checklist, an explicit user-approval validation checkpoint (step 3 gate with AskUserQuestion), and explicit feedback loops (hook failure → fix and re-commit; 'Back findings with sources' → re-present → return to gate; missing plan file → go back and write it). | 5 / 5 |
Progressive Disclosure | Five real reference files, verified one level deep (only intra-set links), each routed by explicit conditions ('for maintainers changing the skill, not for running it'; platform precedence selects github.md/gitlab.md; local mode uses local.md), with the body retaining only the workflow. | 5 / 5 |
Total | 19 / 20 Passed |