aaronartistzhang-afk

audience-brief

Use when the user wants to package a finding, change, status, or plan into a ready-to-paste Feishu/Lark message aimed at a specific audience — RD (研发), 运营 (ops), 老板 (leadership/boss), or a local team. Produces the right register, language, length, and evidence density per audience, then routes ops/boss drafts through the humanizer skill and stops at a preview for the user to confirm before they send. Triggers: "整理一段话我发给研发", "给运营的人话版", "发老板", "package this as an RD message", "发飞书群", "share 给运营", "我发研发用", "给 local team 的版本".

blind-ab-verify

防污染盲测 A/B 闭环 —— 改动落地后，两个变体谁更好且需要人工盲读时，用隔离分臂子代理生成、盲包组装（含阳性/阴性对照 + do-no-harm 硬门）、四栏盲评、回灌解析、三段式 verdict，并强��声明保真度天花板（prompt-level ≠ live pipeline）。当用户说「A/B 验证这个改动」「盲测一下」「打出来我盲读」「防止上下文污染」「两版哪个好」「/blind-ab-verify」，或一个 skill/prompt 改动已落地、需要在不被变体标签污染的前提下判两版优劣时使用。不适用于：动工前的对抗式方案设计 / 审查（用 debate 或 codex-review-gate，属左移）；单轮无对照的主观比稿；需要真实线上流量的效果实验（那要 eng-gated 的 live-RPC）。单版复现：「帮我盲跑一遍这个 skill 看说明书全不全」「干净目录独立复现」「skill 打包验收」 "blind-repro this skill package" → 走模式 B（单版说明书完备性、非两版对比）。

change-triage

Use when the user has to decide which of several candidate changes are worth doing, or when this session is about to propose two or more optional changes and needs the user to pick scope, cost, or priority. Decision triage / cost-benefit / backlog prioritization: force every candidate into a fixed table whose first column is the concrete harm of NOT doing it. 触发词：「不修会怎样」「做的必要性是什么」「值不值得做」「列举 backlog 每项必要性/风险」「这几个决策点我无法决策」“先别送了，你先说明一下你今天都在改什么”“这些候选改动哪个先做”。English triggers: “is this worth doing”, “what breaks if we don't”, “necessity of this change”, “triage my backlog”, “help me decide scope/priority”. 不适用：审对不对（use a review gate）、单个明显必修的 bug（直接修）、纯排序没问代价（“哪个先做”无候选上下文）。

codex-review-gate

跨模型只读审查门禁 —— 用 codex（GPT-5.6 Sol, xhigh reasoning）在只读沙箱里审查方案 / diff / SQL / 对外数字，提取 GO/NO-GO 裁决并循环到收敛，以及编排「老规矩」全流程（Fable 出 plan → plan 门禁 → Opus 子代理落地 → diff 门禁 → 测试）。当用户说「老规矩」「让 codex review 一下」「上集群前记得审查这个 SQL」「这些数字对外前复算一下」「跑一遍门禁」「/codex-review-gate」，或在方案完成 / 实现完成 / SQL 上集群前 / 结论数字进对外文档前需要一道独立审查关卡时使用。不适用于：产出 PRD（用 debate）、通用代码评审纪律（用 requesting/receiving-code-review）。

lark-comment-loop

Use when 处理飞书文档（docx）里的未解决评论，尤其是"只看今天的评论"「拉今天的评论逐条改」「跑一轮评论迭代」「把 PRD 评论闭环」「回复并解决文档评论」这类需求。覆盖 PRD 迭代线（改完 resolve）与数据报告线（只回复不 resolve）两套相反规矩。飞书评论 API 无服务端时间过滤，本 skill 负责客户端按 create_time 过滤 + 归类 + propose-first 落改 + 逐条回复。触发词：「文档评论」「拉未解决评论」「只看今天的评论」「评论迭代」「resolve 评论」「lark-comment-loop」。

lark-peer-feedback-drafting

Use when a person who holds a FORMAL evaluation/feedback responsibility (inside an HR process) wants to draft an evidence-based, first-cut peer-feedback write-up from their OWN visible Lark/Feishu collaboration records. The skill uses lark-cli to assemble first-hand facts (1:1 work chat + a user-confirmed whitelist of groups), builds a re-checkable evidence ledger, and drafts a fact-based feedback draft with a tiered retrieval budget. Human-in-the-loop throughout: the human sets the rating, the human finalizes the wording, and the human submits. NOT for covert investigation, disciplinary evidence-gathering, digging up dirt, surveillance, or evaluating sensitive/protected attributes. 触发词：「写同事评价初稿」「基于协作记录起草反馈」「peer feedback 初稿」「360 评价取证初稿」「lark-peer-feedback-drafting」.

metric-change-attribution

Use when a metric moved between two periods (week-over-week, month-over-month, pre/post, A-vs-B) and you need to attribute WHY — which dimension members, segments, or factors drove it, ranked and drilled to root cause. Covers ratio/rate metrics (reach rate, CTR, conversion, retention, win rate, ARPU) and additive totals (revenue, events, DAU). Triggers: "why did X change", "what drove the drop/spike", "归因", "解释度", "拆解 WoW/MoM 变化", "which segment caused the change", "attribute the change".

session-handoff

Use when work must be handed from this AI coding session to another session on the same machine, or to a later session of the same person. Session handoff / context handoff: produce a fixed three-line header (state / what you do next / what you decide) plus a six-section HANDOFF.md so the receiver sees 'fixed or not fixed?' at a glance. 触发词：「另一个 session」「转给那个 session」「你把这个发给 XX session」「交接」“收工交接”“写个交接”「HANDOFF」。English triggers: “hand this off to the other session”, “write a handoff”, “end-of-shift handoff”, “continue this in my next session”. 不适用：派子代理/并行任务（parallel-agent dispatch）、给人发消息（audience-tailored message）、会话超时/重连（不是交接）。

agent-fleet-guard

Use before launching any large agent orchestration (fan-out) — roughly ≥20 subagents or an estimated ≥1M subagent tokens — and while such a fleet is in flight. Covers: quoting the cost to the human before dispatch, refusing to launch during unstable API connections, in-flight budget stop-loss, T+60s batch health checks, resume-cache traps, and two shell pitfalls that silently break wait loops. 触发词：大编排 / 批量派代理 / fan-out / 20 个代理 / 先报价 / 连接不稳 / 会不会烧太多 token / 这批代理要不要报价。English: “quote this fan-out first”, “is the fleet healthy”, “we're burning tokens”. Not for dispatching one or two agents — that needs no gate.

skill-evolve-loop

Use when sedimenting a lesson from a real run into an existing skill, or when a skill was just edited and needs a post-change regression before it ships. Post-edit regression / skill evolution: locate the smallest section to change, propose the diff first, then run five classes of behavior tests (routing / regression / discipline / content / gate-under-pressure) via parallel subagents, and backfill only what failed. 触发词：沉淀到 skill 里边 / 把这轮教训写进 skill / skill 更新了跑一轮验证 / 改后回归 / 五类行为测试。English: “fold this lesson into the skill”, “post-change regression”, “did my skill edit break anything”. Not for writing a brand-new skill from scratch, and not for blind-repro acceptance of a packaged skill (that's a blind, isolated reproducer's job).

meeting-notes

飞书/Lark 妙记 URL → 精炼会议纪要带 @ 责任人 → 发到群或 DM。当用户提到"总结会议纪要 / 总结妙记 / meeting notes / 飞书会议总结 / 妙记总结 / minute summary / 发到群" 等触发词，且消息里包含 https://*.larkoffice.com/minutes/ 或 https://*.feishu.cn/minutes/ URL 时使用。支持中文/英文/双语三种输出，支持 dry-run 预览，支持 P2P 自送。