Content
67%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 strongly structured, highly actionable instruction skill with a genuinely sequenced workflow and honest limits (Refusals, Known limitations). The main quality cost is redundancy: the pollution-isolation principle is restated so many times that it inflates token cost without adding guidance, and the analysis steps themselves rely on judgment more than the surrounding scaffolding.
Suggestions
Consolidate the persona/rubric isolation rationale into one section (e.g. keep '⚠️ 污染隔离' and reduce the rest to one-line pointers in Key Rules / Integration) — the same leak argument is currently made five times.
Add a brief post-write verification step in Phase 4/5 (check audience.md contains header fields, verified/hypothesized/anti-persona sections per the template) to close the workflow's validation gap.
Tighten Phase 2 with concrete operational hints (e.g. how to bucket a comment into 自我认同/情绪寄存/反驳点/语言, or what counts as one evidence comment) so the analytical core matches the specificity of the surrounding phases.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational (no teaching of known concepts), but the persona/rubric isolation rationale is explained at length in at least three dedicated places ('核信定位 — 绝不混', '⚠️ 污染隔离', Key Rule 4) plus two Refusals and the Integration section — noticeably more than the 'minor instances' of anchor 4. Fits anchor 3: could be tightened. | 3 / 5 |
Actionability | Concrete and executable for an instruction skill: named constants (MIN_RETROS_FOR_DATA_GROUNDED=3, MIN_COMMENTS_PER_TRAIT=3, SEED_FROM_BENCHMARK=auto), explicit file paths, a Phase 1 decision table, evidence rules requiring 出处 and counts, a template link, and a full example output report. Falls short of anchor 5 because the core analysis mechanics (how to cluster comments, '统计占比') are left entirely to judgment. | 4 / 5 |
Workflow Clarity | Six phases are clearly sequenced with an overview diagram and explicit checkpoints (Phase 1 data-sufficiency gate with exit, Phase 2 evidence-count gate, Phase 3 cross-check flag, Phase 5 diff vs previous version). The destructive-op cap does not apply — the overwrite rebuild preserves version history and is explicitly documented as re-runnable. Minor gap versus anchor 5: no post-write verification that audience.md matches the template structure. | 4 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ absent), so the skill is a single well-sectioned file with one-level-deep references (templates/audience.template.md, ../cheat-score-blind/SKILL.md). Good structure and navigation throughout; falls short of anchor 5 because the referenced template and sibling-skill paths live outside the skill bundle and their content is not summarized inline as a fallback. | 4 / 5 |
Total | 15 / 20 Passed |