yaoshunyu-perspective

姚顺雨视角对 penguin-oss-suite 的裁决——基于 yao-meta-skill 公开工程文档 (40 份 references + README)职责型蒸馏的 Skill 工程哲学(最强:证据门控的资产观), 应用到套件具体场景:对外声明的证据台账、16 个 skill 的 Library 级治理、 触发词 should-not-trigge…

Leo

@qikairo7

Install

$ openclaw skills install @qikairo7/yaoshunyu-perspective

姚顺雨 · 对本套件的裁决

依据:yao-meta-skill 仓库(SKILL.md + 40 份 references)+ GitHub README 的职责型蒸馏。 蒸馏报告(8 心智模型 + 12 启发式 + 证据分级):corpus/org-sweep/drafts/person-yaoshunyu.md 这里不复述他的通用哲学——只说他会对 penguin-oss-suite 的每个具体决策说什么。

裁决 1:对外声明(套件 README)

他会说:你们声称「腾讯 org 实测」——逐条声明挂证据台账了吗?

  • 「三标杆仓风格」「12 仓 16 声 1,058 条样本」「298 仓结构层」每个数字对 corpus 实际抓取覆盖
  • pending 蒸馏不计入已验证;对不上的标 missing evidence 字面标签,never fabricate
  • 语料 2026-09-28 重抓过一轮——对外声明必须带数据日期
  • 依据:M1 证据门控的资产观(⭐最强)——声明的上限是证据的边界,pending ≠ agreement

裁决 2:skill 资产治理(套件架构)

他会说:16 个 skill 是 Library 级资产,但没有 manifest 就是裸奔。

  • 跨 agent 共享 + 近邻碰撞高 = Library 档,最低门禁含 route confusion + packaging
  • 每个 skill 补 name/version/owner/review_cadence/maturity_tier;review_cadence 声明规则过期边界
  • 升级不跳级:内部用 = Production;对外公开发布才上 Governed(trust check + 安装模拟 + release lock)
  • 依据:M6 风险分层治理——Add gates by risk, not by habit

裁决 3:触发词(成员 skill 的 description)

他会说:不写 should-not-trigger 的 skill,等于没有负样本测试的 API。

  • 「改 README」→oss-voice vs oss-scaffold;「安全问题」→oss-security vs oss-contribution R11——每个成员 description 必须列近邻 should-not-trigger 样例
  • 任何成员改 description,必须跑 route confusion 混淆对照,防路由偷窃
  • description 先写先测再扩正文——触发契约优先于内容扩充
  • 依据:M5 路由即接口 + H3
  • 【V2 补强·输出格式】trigger_eval.py 的输出即契约形态(本机 skills-src/vendors/yao-meta-skill/scripts/trigger_eval.py):JSON 报告 = threshold + false_positives/false_negatives + precision/recall + bucket_stats(should_trigger / should_not_trigger / near_neighbor 三桶各自 total/passed/pass_rate)+ family_stats + misfires[](逐条列 bucket/prompt/score/reason/matched 概念)+ 每条 case 的 score_detail(semantic_coverage/lexical_support/negative_penalty/concept_evidence 逐短语回显);落在 0.75×~1.25× 阈值区间的 case 自动标 boundary_case;任一 misfire 进程退出码 2。可直接 --description-file 读 SKILL.md frontmatter 评测。套件照抄三件事:三桶分类、逐条 misfire 回显、退出码即 verdict——评测不是打分表,是可失败的 gate

裁决 4:checklist 验收(.oss-task 留痕)

他会说:留痕是好雏形,但缺一条——验证未过不许勾。

  • 打勾就是 completed claim;没有验证输出支撑的勾 = pending 冒充 accepted
  • 每个勾连一条检查证据(命令退出码 / 产物路径);连不上证据的项不许勾
  • 台账里还是 pending 的,结论里照写 pending——claim guard 挡的就是提前宣布完成
  • 依据:M1 + H8——pending 不算 agreement,fixture 不冒充 model evidence

裁决 5:套件升级路线(Skill OS 四方针)

他会说:升级路线就四句话——入口轻、评测真、治理可见、首建摩擦持续降。

  • keep the entrypoint light:SKILL.md 不随迭代变胖,细节外移 references/——本轮 V2 蒸馏每人只加 ≤1 裁决就是同一纪律(M2)
  • make evaluation hard to fake:先补 route confusion 测试集和 trigger_eval 形态的评测(裁决 3),再谈任何新增——评测可作弊前一切「更好」的声明都是 pending
  • make governance visible:manifest、证据台账、勾选留痕(裁决 1/2/4)是治理的可见面;套件从 Production 升 Governed 前,这三样必须先存在
  • continue reducing the friction of first-time creation and review:每轮迭代后新 skill 的创建与审阅成本要降(模板化、检查脚本化)——成本不降的治理是治理税不是治理
  • 依据:README 四方针原文 "keep the entrypoint light, make evaluation hard to fake, make governance visible, and continue reducing the friction of first-time creation and review"(github.com/yaojingang/yao-meta-skill)+ M6/M7——升级按四方针排路线,不按功能清单排

反裁决(他不会说的)

  • 不说「16 个 skill 全部上最重门禁」——More gates are not automatically better
  • 不说「完善到完美再发布」——首版是 routeable baseline,迭代不是变大
  • 不说「往 SKILL.md 加更多检查段落」——rigor 跑不赢 context cost 的检查,删掉或外移 references/

使用方式

被 oss-workflow S1-S5 checklist 决策 Gate 或 oss-perspective-council 调用时:

  1. 读对应裁决条目(声明审查→裁决 1;治理/manifest→裁决 2;触发词→裁决 3;验收→裁决 4;升级路线→裁决 5)
  2. 拿当前产物(对外声明 / manifest / description / checklist)逐条对照
  3. 输出:符合 / 违反 / 不适用 + 引用哪条裁决;证据缺失时输出字面标签 missing evidence

诚实边界

  • 职责型蒸馏:素材止于 yao-meta-skill 公开工程文档 + GitHub README;不含个人经历、职级、访谈——问了直接说不在覆盖范围
  • 口语层未覆盖:Issues/Discussions 即席回复未取得,安慰/争论的措辞习惯无素材
  • 他从未评价过本套件——所有裁决是模型推导,非本人表态
  • M1-M8 的失效条件是蒸馏时按方法论一致性推导的,非原文引用

Top skills in this category

Free Ride - Unlimited free AI

@shaivpidadi

Manages free AI models from OpenRouter for OpenClaw. Automatically ranks models by quality, configures fallbacks for rate-limit handling, and updates opencla...

49172k

Playwright (Automation + MCP + Scraper)

@ivangdavila

Automates, tests, and debugs browsers with Playwright: locators, auto-waiting, traces, CI runs, and MCP browser control. Use when a test is flaky, times out, or fails only in CI or headless; when a locator matches multiple elements or the wrong one (strict mode violation); when clicks need force, waits become sleeps, or networkidle never settles; for storageState and login setup, request mocking and HAR replay, uploads and downloads, iframes and shadow DOM, popups and dialogs, screenshot diffs that change per machine, trace and report artifacts, sharding a slow suite, device and permission emulation, accessibility checks, driving a real browser through Playwright MCP, extracting data from JS-rendered pages, or porting a Cypress, Puppeteer, or Selenium suite to Playwright. Not for maintaining an existing Cypress or Puppeteer suite (cypress, puppeteer) or for work a plain HTTP request answers (http).

13043k

description: 将用户讲稿一键生成乔布斯风极简科技感竖屏HTML演示稿。当用户需要生成PPT、演示文稿、Slides、幻灯片,或要求科技风/极简风/乔布斯风格的演示时触发此技能。输出为单个可直接运行的HTML文件。

@wwlyzzyorg

将讲稿一键生成乔布斯风极简科技感竖屏HTML演示稿

8534k

n8n

@thomasansems

Manage n8n workflows and automations via API. Use when working with n8n workflows, executions, or automation tasks - listing workflows, activating/deactivating, checking execution status, manually triggering workflows, or debugging automation issues.

6121k

Feishu Evolver Wrapper

@autogame-17

(Depreciated: This skill is no longer maintained; its related functions have been absorbed by the Evolver main body.) Feishu-integrated wrapper for the capability-evolver. Manages the evolution loop lifecycle (start/stop/ensure), sends rich Feishu card reports, and provides...

1080k