Content
73%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-sequenced review workflow with copy-paste commands, explicit validation checkpoints, and a real, well-signaled one-level reference bundle. Its main weakness is token efficiency: the marker-resolution and gate contract is duplicated across multiple body sections instead of being delegated once to the existing CHECK-MODE.md reference.
Suggestions
Collapse the duplicated marker/gate contract ('Machine-readable gate result', 'Resolving markers before the gate', and the Output Format validation notes) into a single short summary plus one pointer to references/CHECK-MODE.md, which already owns the full procedure.
Move the generic 'Runtime-aware questions' and 'Shared project skill-context' sections into a shared project-context reference or the skill-context layer, keeping the body focused on the review workflow itself.
Trim rationale prose (e.g. the 'Why: Opus 4.8…' note and 'The dispatch is paid only in runs that…' paragraph) or relocate it to the relevant reference file; the body needs the rule, not the design history.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~430-line body restates the marker/validation/gate contract across at least three sections ("Behavior", "Machine-readable gate result", "Resolving markers before the gate", and the Output Format notes), and inlines generic harness-contract text ("Runtime-aware questions", "Shared project skill-context") plus rationale prose ("The dispatch is paid only in runs that…") that belongs in references/CHECK-MODE.md — noticeably verbose with several padded/duplicated sections, below the 'mostly efficient' bar of 3. | 2 / 5 |
Actionability | Guidance is fully executable throughout: exact commands ("git diff --cached", "git rev-parse --verify <argument>", "git log --oneline --reverse <ref>..HEAD", "git show <commit-hash>"), ready-to-emit AskUserQuestion prompts with options, concrete flag-routing examples ("+check +coupling 123"), and complete copy-paste JSON gate blocks including a worked failure example — matching the 5 anchor's coverage of common cases. | 5 / 5 |
Workflow Clarity | The routing chain (empty → staged, PR URL/digits → PR, validate ref → commits or ask), per-mode numbered steps, commit-count checkpoint, freshness checks, and the validate-then-recompute marker lifecycle form a clear sequence with explicit validation steps, error-recovery paths (re-validate a corrected ref, re-run /aif-review +check), and checklists — the 5 anchor. The review is read-only, so the destructive/batch cap does not apply. | 5 / 5 |
Progressive Disclosure | References are one level deep, clearly signaled with purpose (PR-SOURCE.md, CHECK-MODE.md, SEVERITY.md, COUPLING-REVIEW.md, API-COMPATIBILITY.md, pr-source.mjs, check-workflow.mjs — all verified present in ./references/), and heavy procedures are appropriately split out. The gap versus a 5: the multi-page marker/gate semantics are inlined in the body while also living in CHECK-MODE.md, so the body is not a lean overview. | 4 / 5 |
Total | 16 / 20 Passed |