Conciseness

The body is telegraphic and every line instructs (e.g. "正文、评论、字幕分别核对是否已拿到；够了就直接回答，不装渠道、不登录") with zero explanation of concepts Claude already knows. It matches anchor 5 (lean, every token earns its place); anchor 4 would require trimmable over-explanation, and the only repetition (stop-rules echoed across error-table rows) keeps each row self-contained for mid-task lookup rather than padding.

Actionability

Fully executable guidance: a copy-paste example (`agent-reach xiaohongshu search "咖啡" --limit 5`), per-error remediation commands (`agent-reach add <渠道>`, `agent-reach configure <key>`, `agent-reach doctor --text`), and a user-intent-to-command mapping table. Anchor 5's 'specific examples cover the common cases' is met; the {{CHANNEL_TABLE}}/{{VERSION}} strings are build-time substitutions and the body routes uncertain cases to `agent-reach channels` for the authoritative action list, leaving no executable gap.

Workflow Clarity

The routing workflow has an explicit verification checkpoint ("正文、评论、字幕分别核对是否已拿到"), the 9-code error table provides feedback loops with explicit stop conditions ("不循环重试", "按 retry_after_seconds 等", "过一会儿…再试"), and destructive operations are gated (`agent-reach uninstall --dry-run` → user confirms → run; update only after consent). This matches anchor 5 (explicit validation steps, feedback loops for error recovery); anchor 4's 'minor validation gaps' is not met.

Progressive Disclosure

The ~105-line body stays an overview while delegating detail to three one-level-deep references, each verified to exist in the bundle and each with an explicit read-trigger: "用户要登录、退出、换账号，或遇到 NEED_LOGIN 时读它" (references/sessions.md), "用户问'能查什么/查得全吗'或要判断结果完整度时读它" (references/capabilities.md), and "更新失败、浏览器起不来或要卸载时读它" (references/maintenance.md). None of the three references points to further files, so navigation is exactly one level deep — anchor 5.

Specificity

Quotes a dozen concrete actions: "只读读取各平台内容：搜索、读帖、看评论、取字幕、查行情和职位" plus management verbs "安装、加上/去掉某个平台、登录/退出、看状态、更新、卸载". This matches the anchor-5 example (multiple specific concrete actions, comprehensive coverage); anchor 4 would require minor gaps, but content fetching, triggering, and lifecycle management are all covered. Voice is third person ("只读读取…", "用户提到…"), so no pronoun penalty applies.

Completeness

Explicitly answers what ("只读读取各平台内容：搜索、读帖、看评论、取字幕、查行情和职位") and when with three concrete trigger scenarios ("Use when 用户提到或贴出链接…", fallback failure, and management requests), plus explicit non-goals ("不做：发帖、评论、点赞、私信等任何写操作"). Clearly anchor 5; anchor 4's weaker 'when' does not apply since triggers are explicit and varied.

Trigger Term Quality

Includes platform synonyms "小红书/xiaohongshu、B站/bilibili、Twitter/X、雪球/股票行情、LinkedIn/Boss直聘", natural user phrasings ("帮我搜/看看/读一下这个链接"), and a fallback condition ("自带网页搜索/读取不可用、失败或拿不到正文"). Coverage with synonyms matches anchor 5; anchor 4 would need meaningful natural terms missing, and none are.

Distinctiveness Conflict Risk

Clear read-only multi-platform niche with distinct triggers, and the explicit precedence rule "公开内容优先用 Agent 自带工具，Reach 补缺口" minimizes conflict with built-in web-search skills. Anchor 5 (clear niche, minimal conflict risk) fits; anchor 4's overlap concern is directly defused by that precedence statement.