Content
63%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.
The body is exceptionally actionable, with real commands, explicit fail-closed validation gates, and well-worked error-recovery paths — top marks on execution guidance. It is severely let down by token efficiency and file structure: the same governance rules are duplicated across many sections, dated incident narratives are embedded in normative text, and large self-contained topics that belong in reference files are all inlined in a ~770-line monolith with no bundle files at all.
Suggestions
Move self-contained topics — 浏览器验证政策, 判例 #4152/#4163, remote review QA (Q1-Q4), and the reviewer downgrade table — into reference files under references/ and keep only a one-line pointer plus the operative rule in SKILL.md.
Consolidate the continuity rules currently restated in Review Provenance Matrix, Evidence Manifest, Evidence Validation Checker, Quick Reference, and Common Mistakes into one canonical section, leaving the others as short pointers; extract the multi-hundred-word 'stale' table cell into its own subsection.
Relocate dated incident history (PR #1300, #160, #2214, F168 opus5 gate 实录, specific main SHAs) into a dedicated '历史判例 / old patterns' section so time-sensitive details stop inflating the normative text.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | At ~770 lines, the body restates the same continuity/provenance rules across five or more sections (Provenance Matrix, Evidence Manifest, Evidence Validation Checker, Quick Reference, Common Mistakes) and weaves dated incident history ('2026-04-20 PR #1300 确认', '2026-09-20 opus5 六次 gate 实录', main SHA 'a3aa1ceded') into normative text, plus multi-hundred-word prose inside single table cells (the 'stale' row). This matches 'noticeably verbose; several unnecessary explanations or padded sections'; not 1 because almost none of it explains concepts Claude already knows — it is genuinely project-specific policy, just far over token budget and heavily duplicated. | 2 / 5 |
Actionability | The guidance is fully executable: real shell commands for PR creation (HEREDOC body), body-handle guards with rg patterns, hotfix detection with jq parsing and fail-closed exits, classify-merge-outcome.mjs with explicit tri-state exit-code case handling, worktree finish/inventory invocations, and worked pass/fail output examples for evidence validation. This matches the 'fully executable; copy-paste ready commands covering common cases' anchor. | 5 / 5 |
Workflow Clarity | The merge pipeline has explicit validation checkpoints (E1-E5 hard conditions with BLOCKED failure actions) and genuine error-recovery feedback loops (re-validate after doc fix, pending-vs-failure merge exit triage, re-run check-feature-truth after post-merge sync), plus checklists. Not 5 because the step sequence is incoherent in places: Step 7.5a executes before Step 7 in the command block, and Steps 0, 5.1, and 6.1 are referenced but never defined anywhere in the document. | 4 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent), so everything — browser verification policy, remote-review QA, precedent case studies #4152/#4163, reviewer downgrade matrices — is inlined in one monolithic SKILL.md. External references (../.cat-cafe-shared-refs/pr-template.md, pr-signals.md, sop-definitions/development.yaml) are clearly signaled, which keeps this at 'some structure but content that should be separate is inline' rather than the 'minimal structure' anchor 2. | 3 / 5 |
Total | 14 / 20 Passed |