Content
77%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 and validation-heavy, with exact commands, JSON contracts, and fail-closed outcomes for every mode, giving it top workflow clarity. Its weaknesses are length-driven: a 683-line monolith that repeats the trusted-check and bilingual rules multiple times and keeps mode-specific detail inline instead of splitting it into reference files.
Suggestions
Split each mode (local working tree, assess-candidates, develop-issue, address-review) into its own reference file under references/, keeping SKILL.md as a routing overview with the shared Rules for Every Mode.
Factor the trusted verification command list (npm run build/typecheck/lint, focused Vitest, bundle integration, generate:settings-schema) into a single stated-once section instead of repeating it verbatim in GitHub Actions Rules, develop-issue step 6, and the address-review finish block.
Trim dated fleet statistics and repeated justifications (e.g. the takeover-fleet self-review numbers and the twice-stated bilingual translation rules) to the operative rule plus a one-line rationale.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Every rule is domain-specific operational policy rather than known-concept padding, but the 683-line body repeats material: the trusted-check command list appears verbatim three times (GitHub Actions Rules, develop-issue step 6, address-review finish), the bilingual-output rule is stated twice, and self-review fleet statistics ("measured on the takeover fleet (40 PRs, 2026-09-10)") pad the rationale. This fits the mostly-efficient-but-could-be-tightened anchor 3, not anchor 2 since no section explains things Claude already knows. | 3 / 5 |
Actionability | The guidance is fully executable: exact shell commands ("env -u SANDBOX QWEN_SANDBOX=true \"${QWEN_CODE_CLI:-qwen}\" review run --approval-mode auto --effort high --json --quiet"), exact JSON output shapes for decision.json, self-review.json, comment-replies.json and test-weakening.json, exact git expressions, file paths, and terminal statuses. It matches the copy-paste-ready anchor 5. | 5 / 5 |
Workflow Clarity | Each mode has a numbered sequence with pervasive explicit validation checkpoints (entry/exit content fingerprints, an exhaustive list of fail-closed completion conditions, mutation probes before commit) and genuine feedback loops (fix-and-rerun while an evidence-backed hypothesis remains, iterate review rounds until convergence with STALLED/BLOCKED/CONVERGED outcomes). This is the clear-sequence-with-validation-and-recovery-loop anchor 5. | 5 / 5 |
Progressive Disclosure | The body has real section structure (per-mode headers, a shared GitHub Actions Rules section) so it is not the minimal-structure anchor 2, but it is a single 683-line monolith where per-mode detail (address-review alone spans ~380 lines), bilingual-output rules, and gate-parsing rules clearly belong in separate reference files; the only referenced files (.qwen/skills/*, AGENTS.md) are external to the bundle, and the bundled scripts/run-agent.mjs is never referenced. This matches the some-structure-but-should-be-split anchor 3. | 3 / 5 |
Total | 16 / 20 Passed |