Content
56%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 body with concrete commands, thresholds, and verifiable executable gates, undermined by substantial redundancy — the same rules are repeated across four different sections with anecdotal padding — and by case-study tables that belong in separate reference files. Cross-reference and numbering errors (nine vs ten 铁律, wrong rule pointers) add minor coherence costs.
Suggestions
State each 铁律 once and demote the 反面案例速查 and 业界对照验证 tables to a references/ file (e.g. references/case-studies.md), keeping only one-line pointers in SKILL.md — this addresses both the conciseness redundancy and the inline-vs-separate-file structure.
Fix the internal inconsistencies: the description/frontmatter says 九条 but there are ten 铁律, 铁律三's numbered steps run 1,2,3,2,3,4,5, and checklist item 9 plus one table row cite 铁律九 for the >300 steps/>500 万 input threshold that actually lives in 铁律十.
Replace the <本skill> and <流程记录.json> placeholders with a concrete skill-directory-relative path and a one-line example of the JSONL/flow-record format so the gate commands are copy-paste ready.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~264-line body restates the same rules in up to four places (铁律 sections, canonical sections, the 回合开始前自查 checklist, and the 20-row 反面案例速查 table) and pads rules with session-anecdote evidence ("来源实证：2026-08-24 … 375 万输入 token、16 次压缩"), so several sections are noticeably redundant rather than tight. | 2 / 5 |
Actionability | Guidance is mostly executable — concrete commands ("python <本skill>/scripts/flow_check.py --check-session <记录.jsonl>"), tool call signatures with parameters ("job_output(job_id, wait=true, timeout_ms=60000)", "queue_cli_request … target_model+effort"), and numeric thresholds (≥3 repeats, >300 lines, ≤5 reads) backed by real bundle scripts. Minor gaps: placeholders like <本skill> and <记录.jsonl> require the operator to resolve paths and input format. | 4 / 5 |
Workflow Clarity | Multi-step processes are clearly sequenced with executable validation gates and feedback loops ("FAIL 则禁止实施", "复跑至绿", "--selftest exit 0 才可交付"), and the 回合开始前自查 acts as an explicit checklist. However, 铁律三's ordered steps break numbering (1, 2, 3, 2, 3, 4, 5) and two cross-references point to the wrong rule ("见铁律九" for the staged-handoff threshold that is actually 铁律十; the description also says 九条 while ten 铁律 exist), creating minor incoherence. | 4 / 5 |
Progressive Disclosure | Both bundle scripts (scripts/flow_check.py, scripts/takeover_check.py) are real and referenced with concrete commands, and section headers are clear, but large supporting material — the 20-row 反面案例速查 table and the 业界对照验证 table — is fully inline in SKILL.md instead of split into reference files, so content that should be separate is inline. | 3 / 5 |
Total | 13 / 20 Passed |