Content
92%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 rigorously engineered operational manual: every step is executable, batch/destructive operations are wrapped in verification and retry loops, and historical precedent is properly split into a one-level reference. The only weakness is occasional inline incident history with specific dates and usernames that already lives in precedents.md.
Suggestions
Move the inline dated incident details (e.g. the 2026-09-27 capacity anecdote, the huangy7/stimQQ and SliceLoft incidents) into references/precedents.md, keeping only the one-line warning plus the 「→ 判例：X」 pointer in the body.
The 硬红线 section intentionally restates rules detailed in sections 3–9; consider compressing the later restatements to pure cross-references to cut the ~26KB body further.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational — no explanations of concepts Claude already knows, and every rule carries decision weight. Minor trimmable material exists: inline dated incident anecdotes (「2026-09-27 实测」「2026-10-21 事故」, huangy7/stimQQ usernames) duplicate what references/precedents.md already archives, which time-sensitive detail penalizes per the guidelines. | 4 / 5 |
Actionability | Fully executable throughout: copy-paste gh/git/jq/python pipelines, exact thresholds (490,000-byte capacity trigger), concrete jq filters, merge-conflict command sequences, and verbatim reply templates for accept/reject cases. | 5 / 5 |
Workflow Clarity | Clearly sequenced 0→9 pipeline with validation checkpoints everywhere: environment precheck that halts on failure, POST→PATCH→GET verification of every comment, merge retries after re-querying mergeable_state, re-scan until signature residue is 0, and exit-1 guard on invalid scan data — feedback loops fully covering the batch/destructive operations, so no cap applies. | 5 / 5 |
Progressive Disclosure | Two-tier structure is explicit and well-signaled: the body holds process and hard constraints with 「→ 判例：X」 markers, and one-level-deep references/precedents.md (sections A–L, all verified to exist) holds the case history, with a standing instruction not to bulk-load it at runtime. Navigation between the two is unambiguous. | 5 / 5 |
Total | 19 / 20 Passed |