蒸馏一个人的认知操作系统的 Agent Skill — 对外汉语教学法 × AI「第零阶级」方法论 × 跨时空连接思维。Distill a person's cognitive OS. Inspired by nuwa-skill.
- Category: AI workflow infrastructure
- Audience: AI builders, creators, independent developers, and small teams that want repeatable local AI workflows.
- Repository status: Public repository. Keep examples, docs, and issues free of credentials, private data, and machine-specific paths.
- Topics: agent-skills, ai-methodology, chinese-teaching, claude-code, claude-skill, cognitive-framework, mental-models, nuwa-skill
- 蒸馏一个人的认知操作系统的 Agent Skill — 对外汉语教学法 × AI「第零阶级」方法论 × 跨时空连接思维。Distill a person's cognitive OS. Inspired by nuwa-skill.
- Package repeatable AI workflows into reusable local assets.
- Reduce one-off prompt work with versioned skills, guardrails, and handoff files.
This repository is part of a broader AI local-workbench operating model: reusable skills, local automation, auditable configuration, and repeatable delivery workflows.
- Primary language: Not classified by GitHub
- Default branch:
master - Visibility:
PUBLIC - License: Other
CHANGELOG.mdCONTRIBUTING.mdLICENSEREADME.mdSECURITY.mdSKILL.mdexamplesreferences
No universal setup command is declared yet. Start by inspecting the repository map above and the project-specific scripts or docs committed with the code.
- Keep real credentials out of the repository. Use local environment files, GitHub repository secrets, or the deployment platform secret manager.
- If a
.env.examplefile exists, treat it as documentation only; never commit filled-in.envfiles. - Before publishing screenshots, demos, or client examples, remove private names, internal paths, account IDs, and API endpoints.
- The
Repository Hygieneworkflow is a lightweight guardrail, not a replacement for product-specific tests.
- README describes the user, business outcome, and operating boundary.
- Setup or preview commands are current and do not rely on private machine state.
- No real secrets, private user data, or machine-local state are tracked.
- Screenshots, demos, or sample outputs are safe to share publicly when the repository is public.
- Product-specific tests or smoke checks are documented before production use.
- Tighten the fastest path from clone to useful demo.
- Add project-specific screenshots, sample outputs, or a short walkthrough where useful.
- Promote repeated manual steps into scripts, tests, or documented workflows.
- Keep security, privacy, and licensing boundaries explicit as the project evolves.
Maintained by Tony Sheng. This README is written as a business-facing handoff: it should help a future collaborator, client, or reviewer understand why the repository exists, how to inspect it, and what must be true before it is reused or shipped.
| 模块 | 帮谁 | 核心 |
|---|---|---|
| 🀄 思维桥梁教学法 | 中文老师 & 中文学习者 | 教思维不教翻译：四层逐字对照（Chinese / Pinyin / Word-by-Word / English）+ 把语言点拉进历史与跨学科纵深的"文化对照配方" |
| 🤖 「第零阶级」AI 方法论 | 想真正用好 AI 的人 | 能力增益而非卸载（"用 AI 让你变轻松=用错了"）+ 自进化系统 8 大设计原则 + 多模型编排 + token 经济学 |
| 🧩 跨时空连接思维 | 任何做思考 / 写作 / 决策的人 | 7 个心智模型 + 12 条决策启发式 + 7 条 voice 规则，把当下任何事拉进历史长河与别的学科 |
外加一套贯穿全局的 Agentic 研究协议：涉及当下事实必须先查证再断言，绝不编造——这是从 nuwa-skill 继承的诚实纪律。
- 跨时空连接 — 没有孤立知识点，只有节点和连边
- 工程化破界 — 一切皆可拆解、优化、重构的系统
- 自进化系统优先 — 最高产出是"造一个能造系统的系统"，进化本质即递归
- 能力增益而非卸载 — "变轻松"是用错的信号
- 迭代胜过规格 — 先跑 v1 看真实结果，卡在规划比做错更糟
- 验证以现实为准 — 只信磁盘 / 真实渲染 / smoke test，不信自报数字
- 杠杆经济学 — 在"放大产出"与"守住约束"间显式权衡
谱系：Tony = 李笑来「重新定义 + 践行」× 万维钢「科学破常识」× 金观涛 / 史学「大历史纵深」× 李小龙 / 禅「集成不立派」× Karpathy / Sutton「工程审美与 scaling 信仰」，再用余华式「人味与反讽」上色。
tony-skill 遵循开放的 Agent Skills 协议，SKILL.md + references/ 即可被任意兼容运行时加载。
Claude Code
# 克隆到你的 skills 目录
git clone https://github.com/shengdabai/tony-skill ~/.claude/skills/tony-skill随后在对话里自然触发即可，例如：
- "用思维桥梁教学法教我『把』字句" → 自动加载教学模块
- "帮我设计一个能自进化的资讯简报系统" → 自动加载 AI 方法论
- "用 Tony 的方式想想：我该不该用 AI 写日记" → 自动加载心智模型 + voice
Cursor / Codex / 其他：把本仓库作为 skill 目录引入，或把 SKILL.md 内容作为 system prompt 注入；references/ 按需加载。
tony-skill/
├── SKILL.md # 核心：路由 + 心智模型摘要 + voice + 边界 + 研究协议
├── README.md / LICENSE # 仓库文档与许可
├── CONTRIBUTING.md / CHANGELOG.md # 迁移指南与版本记录
├── references/
│ ├── teaching-method.md # 思维桥梁教学法全文（四层对照 + 文化纵深五步配方）
│ ├── ai-method.md # 「第零阶级」playbook（8 原则 + 多模型编排 + token 经济学）
│ ├── mental-models.md # 7 心智模型 + 12 启发式 + 反模式 + 价值观（全证据版）
│ ├── genealogy-and-voice.md # 思想谱系（225 本知识食谱聚类）+ 7 条 voice 规则
│ └── research-protocol.md # Agentic 研究协议 + 诚实边界
└── examples/
├── teaching-example.md # 老师视角备课：「给我打电话」教成思维跃迁
├── learner-self-study-example.md # 学习者视角自助：用四层对照攻克「把」字句
├── ai-leverage-example.md # 自进化系统设计（含一次真实翻车→诊断→修正→落盘）
└── thinking-example.md # Tony voice 思考：该不该用 AI 写日记
遵循 Claude skill-creator 的结构规范，借鉴 nuwa-skill 的"并行研究 → 三重验证 → 技能构建 → 质量测试"四阶段框架。素材由多个 AI agent 并行分析以下私有源蒸馏而成（已移除密钥、私有链接、第三方姓名等直接敏感信息；但作者有意公开了自己的职业身份、知识食谱规模等可画像信息，这是一个自觉的隐私权衡，并非"零信息暴露"）：
- 跨会话的 AI 协作记忆与决策记录
- 225 本 NotebookLM 知识库构成的"知识食谱"（用于反推思想谱系与表达 DNA）
- 真实教学实例与方法论档案
每个心智模型都经过"跨领域复现 / 生成力 / 独特性"三重验证，并诚实标注局限——这套思维擅长生成洞察、驱动产出，但普遍缺"何时该停"的内建判据，文档对此如实交代。
- 教学法的文化对照必须基于真实史料；编一个玄乎的"文化解释"比不讲更糟。
- voice 部分是合理推断（从"读什么"+ 真实教学实例反推），模仿时继承 Tony"多空对冲、自己当裁判"的姿态，不替他对争议选边。
- 涉及当下事实，先查证再说，绝不用"Tony 大概会这么想"掩盖一个本该核实的事实。
tony-skill distills the cognitive operating system of Tony — a career teacher of Chinese as a foreign language (CELTA + TESOL certified, thousands of students taught, per the author) who is also a heavy AI power-user — into one executable Agent Skill.
It captures not what he said, but how he thinks, teaches, and uses AI: transferable mental models, a reusable teaching method, and a copy-able methodology. Install it and your AI will dissect problems, design Chinese-learning, build self-evolving systems, and write in Tony's voice.
This is not role-play. Don't imitate a persona — install a way of thinking. The test is always: does the answer make the other person see deeper, do more, grow — not "does it sound like Tony."
| Module | For whom | Core |
|---|---|---|
| 🀄 The "Thinking Bridge" teaching method | Chinese teachers & learners | Teach thinking, not translation: a four-line gloss (Chinese / Pinyin / Word-by-Word / English) + a recipe for pulling any language point into historical & cross-disciplinary depth |
| 🤖 "Class Zero" AI methodology | Anyone who wants to truly master AI | Augmentation over offloading ("if AI makes you easier, you're using it wrong") + 8 design principles for self-evolving systems + multi-model orchestration + prompt-cache economics |
| 🧩 Cross-temporal connective thinking | Anyone who thinks / writes / decides | 7 mental models + 12 decision heuristics + 7 voice rules that pull any topic into deep history and other fields |
Plus a global agentic research protocol: any present-day fact must be verified before asserted — never fabricated. This honesty discipline is inherited from nuwa-skill.
- Cross-temporal bridging — no isolated facts, only nodes and edges
- Engineering-lens universalism — everything is a system to be re-architected
- Recursive self-improving systems — the highest output is "a system that builds systems"; evolution is recursion
- Augmentation over offloading — "feeling easier" is the signal you're doing it wrong
- Iteration over specification — ship a real v1, getting stuck in planning is worse than being wrong
- Reality-grounded verification — trust disk / rendered output / smoke tests, never self-reported numbers
- Constraint-aware leverage — every decision weighs "amplify output" against "respect the constraint"
tony-skill follows the open Agent Skills protocol — SKILL.md + references/ load into any compatible runtime.
Claude Code
git clone https://github.com/shengdabai/tony-skill ~/.claude/skills/tony-skillThen trigger naturally in chat, e.g.:
- "Teach me the 把 construction with the thinking-bridge method" → loads the teaching module
- "Design a self-evolving news-brief system for me" → loads the AI methodology
- "Think about this the way Tony would: should I use AI to write my journal?" → loads the mental models + voice
Cursor / Codex / others: add this repo as a skill directory, or inject SKILL.md as a system prompt and load references/ on demand.
Built following Anthropic's skill-creator structure conventions, using nuwa-skill's four-phase framework (parallel research → triple validation → skill construction → quality testing). The source material was distilled by multiple AI agents working in parallel over private sources, with secrets, private links, and third-party names removed. (Note: the author intentionally discloses his own professional identity and the scale of his knowledge diet — a deliberate privacy trade-off, not a claim of zero information exposure.):
- cross-session AI-collaboration memory and decision logs
- a "knowledge diet" of 225 NotebookLM notebooks (used to reconstruct intellectual genealogy & expression DNA)
- real teaching cases and methodology archives
Every mental model passes a triple test (cross-domain reproduction / generative power / uniqueness) and honestly states its limits — these frameworks excel at generating insight and driving output, but generally lack a built-in sense of when to stop, which the docs disclose plainly.
- The teaching method's cultural depth must rest on real sources; inventing a mystical "cultural explanation" is worse than not giving one.
- The voice profile is reasoned inference (reverse-engineered from "what he reads" + real teaching cases); imitate his hedge-both-sides, be-your-own-judge stance, don't take sides for him.
- For present-day facts, research first — never let "Tony would probably think…" paper over something that should be checked.
- No built-in stop criteria: the frameworks here excel at generating insight and driving output, but don't tell you when enough is enough. Add your own exit conditions when adapting them to loops or automated pipelines.
- Token-cache economics are provider-dependent: the prompt-cache figures in
references/ai-method.mdapply only when your provider supports prompt caching (e.g. Claude, some OpenAI models). Don't assume the numbers hold on every runtime. - No automated CI: there are no tests validating that routing table entries in
SKILL.mdmatch actual files inreferences/. If you fork and rename files, updateSKILL.mdmanually.
最自然的 fork 方式是用同一套方法蒸馏你自己，做出 your-name-skill。
迁移流程、PII 脱敏清单与诚实边界要求见 CONTRIBUTING.md；版本演进见 CHANGELOG.md。
The most natural fork is to distill yourself with the same method. See CONTRIBUTING.md for the migration workflow, PII checklist, and honesty-boundary requirements.
本项目的灵感、方法论框架与"蒸馏认知而非角色扮演"的核心理念，全部来自
alchaincyf/nuwa-skill —— 花叔（Huashu）的开源杰作。
nuwa-skill 提出了"何必只蒸馏同事"的精彩命题，并设计了并行研究 + 三重验证的造技能流程。
tony-skill 把这套框架从"公众人物的公开资料"迁移到"一个具体的人的私有记忆与知识食谱"，并补上了严格的 PII 脱敏与诚实边界。
This project's inspiration, methodological framework, and the central idea of
distilling cognition rather than performing a persona all come from
alchaincyf/nuwa-skill by Huashu (花叔).
tony-skill adapts that framework from "a public figure's public record" to
"one specific person's private memory and knowledge diet," adding strict PII
de-identification and explicit honesty boundaries.
Also built on Anthropic's skill-creator conformance standard.
MIT © 2026 Tony · tony-skill