Conciseness

The body is dense and imperative with no padding of concepts Claude already knows — the Task prompt template, grep self-check, and worked composite example all earn their tokens. It falls short of 5 because the "不写文件/不预测" constraint is restated across Step 6 ("❌ 写任何文件"), Key Rule 5 ("不写文件。这是 score 与 predict 的核心区别"), and Refusals, which could be consolidated.

Actionability

Fully executable guidance: a copy-paste Task prompt template ("Spawn cheat-score-blind sub-agent... 不要询问用户 —— 你没有用户"), an executable self-check command (`echo "<prompt>" | grep -Ei '播放|阅读|点赞|评论数|实际|retro|复盘|实绩|w$|万$'`), the exact composite formula with a worked calculation, and a concrete output template with example per-dimension reasons. Not 4 because no gaps remain — even refusal-code handling is enumerated.

Workflow Clarity

Six clearly sequenced steps with explicit validation checkpoints: Step 1 stops on missing `.cheat-state.json`/draft/rubric, Step 2 refuses to guess a malformed formula ("不要自己猜"), Step 3 mandates a grep self-check before delegation, and Step 4 defines error-recovery loops for every refusal code (`blocked_contaminated_input` → resend, `rubric_unparseable` → alert user). This matches the anchor with feedback loops and explicit checkpoints; the operation is read-only, so no destructive-op cap applies.

Progressive Disclosure

The body is a clear overview with well-signaled, one-level-deep references — the blind sub-agent's isolation protocol is delegated to [skills/cheat-score-blind/SKILL.md](../cheat-score-blind/SKILL.md) with the specific section ("主 Claude 调用契约"段) named, and section headers (Overview, Constants, Inputs, Workflow, Key Rules, Refusals, Integration) make navigation easy. No bundle files exist to inline, and content is appropriately split rather than buried.

Specificity

Phrases like "给单篇稿子打 rubric 分" and "只在控制台输出，不写文件，不预测" name the domain and 1–2 concrete behaviors, but the description omits what the body actually does: per-dimension 0–5 integer scoring with reasons, composite calculation, and bucket placement. Not 4 because it does not list several specific actions — the constraints stated are scope limits rather than capabilities.

Completeness

Both questions are answered explicitly: what ("给单篇稿子打 rubric 分。只在控制台输出，不写文件，不预测") and when ("触发词：'打分这篇 [path]'/'score this [path]'/…"), plus the relationship to the sibling skill ("是 cheat-predict 之前的轻量探索动作"). Not 4 because the 'when' is already maximally explicit with concrete trigger phrases, not merely implied.

Trigger Term Quality

Four natural trigger phrases in both languages are quoted verbatim — "打分这篇 [path]", "score this [path]", "给这稿子打分", "先打分看看" — covering synonyms a user would actually say. This matches the comprehensive-synonyms anchor; not 4 because no common variation is missing.

Distinctiveness Conflict Risk

The niche is mostly distinct: the "不预测" boundary and "cheat-predict 之前的轻量探索" positioning separate it from the closely related cheat-predict skill. Not 5 because a user asking to "打分" while actually wanting a full prediction is a plausible overlap with cheat-predict, and "先打分看看" is generic enough to collide with other scoring skills in the suite; not 3 because the explicit predict/score split keeps conflict risk minor.