Content
70%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.
An exceptionally actionable, rigorously sequenced operational skill whose workflows leave nothing ambiguous — every command, schema, threshold, and stop condition is spelled out. Its weaknesses are the opposite trade-off: a ~680-line monolithic body with repeated command lists, duplicated bilingual rules, and over-detailed gate internals, and no use of reference files to split the four modes' detail out of the top-level SKILL.md.
Suggestions
State the trusted verification command list once in GitHub Actions Rules and reference it from develop-issue and address-review ("run the trusted checks per GitHub Actions Rules") instead of repeating the full build/typecheck/lint/Vitest/bundle/schema list three times.
Split each mode's detail (address-review's gate rules, the in-round self-review procedure) into per-mode reference files under references/ loaded when the Invocation block names the mode, keeping SKILL.md a lean routing overview with a short summary of each mode.
Consolidate the two bilingual-output rule blocks into one section covering both the PR-comment files (with the details block) and failure.zh.md (plain Markdown, no HTML), and trim gate-internal mechanics (test-skip spellings, byte-truncation behavior) to the decision-relevant rule.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~680-line body is noticeably verbose: the trusted verification-command list (build/typecheck/lint/focused Vitest/bundle/generate:settings-schema) is repeated verbatim in three places, the bilingual-output rules are explained in two long overlapping blocks, and external gate internals (every .skip spelling, byte-truncation mechanics) are over-explained. It avoids anchor 1 only because nothing teaches concepts Claude already knows — it is dense operational content with systemic redundancy rather than pedagogical padding. | 2 / 5 |
Actionability | Guidance is fully executable throughout: exact shell commands (the env -u SANDBOX review run invocation, the self-review CLI command), exact JSON output schemas (decision.json, self-review.json, comment-replies.json, test-weakening.json, deferred-findings.json, growth-audit.json), exact git expressions, and precise numeric thresholds (150 lines, 75 minutes, ~8 findings per round). This matches the anchor-5 example of copy-paste-ready commands covering the common cases. | 5 / 5 |
Workflow Clarity | Every mode is a numbered, explicitly sequenced procedure with fail-closed validation checkpoints (entry/exit fingerprints, terminal-status field checks, event classification), feedback loops (fix-and-rerun failing checks, serial review rounds with progress tests), explicit stop conditions (NO_CHANGES/CONVERGED/BLOCKED/STALLED), and exhaustive checklists. This is a clear anchor-5 match; anchor 4 would imply missing checkpoints, and none are missing. | 5 / 5 |
Progressive Disclosure | The body has real, navigable structure (per-mode ## headers, ### subsections, fenced command and JSON blocks), so it is above anchor 2's "minimal structure". But no references/ or assets/ bundle files exist and everything — including the ~350-line address-review section with its gate rules and the in-round self-review subsection — is inlined in one monolithic SKILL.md, fitting anchor 3's "content that should be separate is inline". | 3 / 5 |
Total | 15 / 20 Passed |