Conciseness

The body is dense and assumes competence — recipes are table rows, tool calls are bare signatures, no elementary concepts are explained — matching "Efficient; minor instances of over-explanation that could be trimmed". It is not 5 because it carries substantial time-sensitive internal history ("2026-05-17 dogfood 实证", "2026-07-08 三刀实测", "46 review P3 补", F085/F195/F196/F169, KD-8, DF-1, SW-1) that is not in a deprecated/old-patterns section and spends tokens an agent doesn't need.

Actionability

Guidance is mostly executable: concrete tool invocations with parameters ("search_evidence(query, hybrid, scope=docs, limit=10)", "graph_resolve(anchor, depth=1, relations=[feature_ref,related_to])"), per-question-type recipes with example queries, and a concrete 5-step worked example — matching "Mostly executable guidance; concrete code or commands with minor gaps". It is not 5 because calls are parameterized patterns rather than copy-paste-ready literals (no fully worked search string per round, and the agent-expansion step is described rather than demonstrated as a runnable command).

Workflow Clarity

Multi-step processes are clearly sequenced with explicit validation checkpoints: the 5-step AUDHD recipe, a per-type stop-criteria table ("≥3 路 + agent expand 二轮 + Read canonical → 无新 anchor 出现 才停", "正反两路 + ≥1 个相关概念都 0 命中即可断言 absent"), and a Common Mistakes table with 错误/后果/修复 that serves as error-recovery feedback — matching "Clear sequence with explicit validation steps; feedback loops for error recovery; checklists for complex processes". Not 4 because checkpoints and recovery guidance are explicit rather than implicit.

Progressive Disclosure

No bundle files exist, and the single SKILL.md is well-organized with clear section headers, each self-contained and navigable — matching "Good structure; most content is appropriately placed; minor organization gaps". It is not 5 because at ~130 dense lines, specialized material (the Ragdoll family calibration section and the F256 Related-directions integration) could be split into reference files, and the "相关" section points at external project docs rather than bundle-relative paths, so references are not all clearly signaled within the skill.

Specificity

The description lists several concrete capabilities — "多刀检索 + recall coverage 策略（8 类题型 recipe）", "多 query / 多 scope union + coverage matrix（item/source/谁提到/直接 vs 间接）+ 一次终局交付" — which matches the anchor "Lists several specific actions; minor gaps in coverage". It falls short of 5 because the extreme compression and internal jargon ("多刀", "Ragdoll家族", "opus 系/fable 系") makes some claimed actions hard to parse as concrete capabilities, and it is not above the anchor since it never pads with vague language.

Completeness

It explicitly answers what ("记忆系统多刀检索 + recall coverage 策略") and when ("Use when: 任务是…") with concrete trigger phrases, plus a "Not for" exclusion list — matching the anchor "Clearly and explicitly answers both what AND when with concrete trigger phrases". Not below 5 because neither half is vague or merely implied.

Trigger Term Quality

The "Use when" clause enumerates the exact natural phrases a user would say — "哪些地方提过 X", "X 的来源 / source map", "有没有提过 Y / absence check", "上次到现在变了什么 / delta", "冷启动 onboard 复杂主题" — covering synonyms in both Chinese and English, matching the anchor "Comprehensive coverage of natural terms including synonyms". It is not below 5 because no common variation of these recall requests is missing.

Distinctiveness Conflict Risk

It carves a clear niche (coverage/multi-pass recall) and explicitly resolves the one real overlap — "GOTCHA: 和 memory-navigation 互补不重叠 — memory-navigation 决定第一刀走哪个工具…本 skill 决定要不要补刀" — plus a "Not for" list steering code-symbol lookups and new-feature work elsewhere, matching "Clear niche with distinct triggers; minimal conflict risk". Not 4 because the overlap risk is not merely minor but actively disambiguated.