Content
85%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A tight, well-sequenced review workflow with an explicit validation gate before commenting. Its one weakness is actionability: the gh command and policy rules are concrete, but the parallel reviewer prompts and tier selection are left abstract.
Suggestions
Add an executable example of the reviewer-subagent prompt or a concrete tier/role mapping so the parallel-review step is copy-paste ready, not just described.
Specify how "lightweight and deep review tiers" map when Stave Auto is unavailable (e.g., concrete role names or prompt templates) instead of deferring to Auto.
Show the exact comment-posting command form (e.g., the `gh pr comment` invocation with inline positioning) for the fallback path, not just "otherwise `gh`".
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Lean and free of concept padding; every line is directive instruction (e.g., "Subagents must not post comments", "Reject anything not clearly real...") and assumes Claude's competence. | 3 / 3 |
Actionability | Provides one concrete command (`gh pr view --json ...`) and specific policy-precedence rules, but the core reviewer-subagent invocations are described abstractly ("Run independent reviewers in parallel", "Give each reviewer...") without executable prompt content or tier definitions. | 2 / 3 |
Workflow Clarity | A clear 5-step sequence with an explicit validation checkpoint in step 4 ("Validate each unique issue with a separate pass before commenting" plus reject rules), giving a feedback loop before the destructive act of posting comments. | 3 / 3 |
Progressive Disclosure | A compact (~40 line), well-organized instruction-only skill with no external bundle files; per the simple-skills note, clear sectioning (Workflow, Fallbacks) earns a 3 without external references. | 3 / 3 |
Total | 11 / 12 Passed |