Content
81%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 orchestration procedure with an unambiguous step sequence, genuine validation feedback loops (checks-red fix-and-retry, no-progress break, full-range re-review), and clean delegation of the record semantics to a referenced rules file. The main costs are repeated invariants that could be consolidated and a loop spec whose abstract operations are defined only via the missing referenced file.
Suggestions
Consolidate the repeated never-green-wash invariant: state it once in Hard rules and have the empty-range abort and flagged-reporting sections reference it by name instead of restating it.
Define or inline the loop's abstract operations (apply_findings, run_fast_checks, open_findings) — even one-line signatures with their return states — so the Step 1 spec is self-contained when findings-bus.md is not at hand.
Merge the STOP_REASON derivation warning (loop comment and Step 2 both explain 'never infer it from the iteration count') into a single place, e.g. the report section only.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence (no explanations of git, CI, or review basics), with every table earning its tokens, but invariants are repeated — the never-green-wash rule appears three times (empty-range abort, 'Surface every flagged finding prominently', Hard rules) and the STOP_REASON warning appears both in the loop comment and Step 2. This matches anchor 4 ('efficient; minor instances... that could be trimmed') rather than 3, since none of the repetition explains concepts Claude already knows. | 4 / 5 |
Actionability | Step 0 gives executable bash (git rev-parse fallback chain, bus path setup), the check discovery is a concrete first-hit-wins list (package.json/Makefile/justfile/Taskfile scripts, then documented lint+typecheck+test targets), the report is a copy-paste template, and the dispatch prompt shape is spelled out. The Step 1 loop itself is control-flow pseudocode with undefined abstractions (apply_findings, run_fast_checks), though harness variance is explicitly justified, so it lands at anchor 4 ('mostly executable... minor gaps') rather than 5. | 4 / 5 |
Workflow Clarity | The sequence is explicit — Step 0 precondition/empty-diff abort, loop A (full-range re-review) → B (apply, one commit per finding) → C (simplify) → D (checks) → Step 2 report — with real feedback loops: red checks trigger 'fix the regression this loop just introduced, then re-run once / if still red: break', a no-progress break exists, and each pass validates the previous iteration's fixes over the full range. STOP_REASON discipline prevents misreported outcomes. This matches anchor 5 (clear sequence, explicit validation, error-recovery loops). | 5 / 5 |
Progressive Disclosure | Good structure with a clearly signaled one-level-deep reference — rules/findings-bus.md is linked with purpose ('Owns the record, the lifecycle, and the convergence predicate. Read it; never restate it') and targeted anchors (#the-lifecycle, #append-only-and-why), and sibling links are peers rather than nesting. However, no bundle directory is provided and the referenced rules/findings-bus.md is absent from the bundle, so its navigation target cannot be verified — anchor 4 ('references mostly clear; minor organization gaps') rather than 5's 'easy navigation'. | 4 / 5 |
Total | 17 / 20 Passed |