Content
66%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 excels at behavioral precision: a gated workflow, anti-fabrication red lines, self-audits, and escape hatches make this persona skill unusually rigorous and directly executable. It is held back by heavy internal duplication — the certainty ladder, signature rules, and key evidence each appear 2-4 times across models, heuristics, and DNA sections — and by an essentially un-navigated 160 KB reference bundle while reference-appropriate content stays inlined in the main file.
Suggestions
Deduplicate systematically: keep the certainty-gradient table and signature/frequency rules in the 表达 DNA section only and have Step 3 reference them; merge 启发式 2/3/4/7 into the corresponding 心智模型 sections (模型 3/2/6/4) since they restate the same evidence and phrases.
Split reference-like material out of SKILL.md into one-level-deep files — e.g. references/timeline.md (人物时间线 + 最新动态) and references/sources.md (附录：调研来源) — keeping the operative role-play rules, workflow, and models in the main file.
Add per-file navigation for the existing research bundle: instead of a bare directory pointer, map each file to its purpose (e.g. '表达 DNA 细节见 references/research/03-expression-dna.md') so the 160 KB of research is discoverable rather than buried.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~640-line body contains systematic duplication: the 5-level certainty gradient table appears twice (Step 3C and 表达 DNA), the synchronous_commit ALL-CAPS example is cited in 模型 2, 4, 6 and 启发式 4, and 启发式 2/3/4/7 largely restate 模型 3/2/6/4; the 人物时间线 and 最新动态 sections also overlap on 2026 events. This goes beyond 'could be tightened' — several whole sections are padded restatements, matching the 'noticeably verbose' anchor rather than the level-3 anchor. | 2 / 5 |
Actionability | For an instruction-only skill the guidance is fully concrete and copy-paste ready: exact openers ("I have to admit I don't have a clear position on this"), a sentence-skeleton catalog (起手/转折/拍板/结尾), named search targets (git.postgresql.org, message-id archives), anchor requirements (commit hash, bug #, message-id), and template phrases. This matches the 'fully executable, specific examples cover common cases' anchor. | 5 / 5 |
Workflow Clarity | The Step 0 speculation gate → Step 1 classification table → Step 2 research protocol (per question type) → Step 3 output recipe is clearly sequenced, with explicit validation checkpoints (anti-fabrication red lines, post-generation certainty-gradient audit, a 5-item self-check command) and genuine feedback loops (escape-hatch rules for fabricated hashes, style slips, and out-of-scope topics with recovery actions). This matches the top anchor: clear sequence, explicit validation, error-recovery loops, and a checklist. | 5 / 5 |
Progressive Disclosure | The bundle contains references/research/ with six large raw files, but the body signals them only once ("调研过程详见 references/research/ 目录") with no per-file navigation or indication of which file covers what. Meanwhile large reference-like content — the full timeline table, 智识谱系, and the sources appendix — is inlined in SKILL.md rather than split out. This matches the anchor 'content that clearly belongs in separate files is inlined; references are buried', and is above level 1 only because the structure is real and the references are one level deep. | 2 / 5 |
Total | 14 / 20 Passed |