A methodology collection that makes AI agents honest, auditable, and dependable — built from real production battles, not theory. 一套让 AI 干活更靠谱的方法论：不编造、不假报完成、不丢记忆。所有方法都来自真实项目实战，不是纸上谈兵。
This repository contains 7 complete methodology skills, all included in this repo — download it and start using them today. Each skill fixes one real problem: 本仓库包含 7 个完整的方法论技能，全部内容都在这里，下载后即可直接使用。每个技能解决一个真实问题：
- The agent says "done" but did nothing. We caught a pipeline where 10,000+ claimed landings were 100% absent — the method described it, the code never did it. AI 说"做完了"但其实什么都没做。 实测抓到过：10,000+ 条声称已落地的记录，实际 0 条存在。
- The agent fabricates data. Full-coverage audits found hallucinated citations, wrong standards, and mis-attributed numbers. AI 编造数据。 全量审计发现过编造的引用、错误的标号、张冠李戴的数字。
- The next agent starts from zero. Project state dies with the conversation. 换个 AI 就失忆。 项目状态跟着对话一起消失，新接手的人从头开始。
Fastest path: open skills/, find the skill that matches your problem, and hand its SKILL.md to your agent (Claude Code, Codex, Hermes, Gemini CLI — any of them). It is the complete methodology, ready to use. Skills with a scripts/ folder (research, rule checks) come with stdlib-only Python scripts — copy them into your project and run.
最快路径：打开 skills/ 目录，找到解决你问题的技能文件夹，把里面的 SKILL.md 整个发给你的 AI（Claude Code / Codex / Hermes / Gemini CLI 都行）——它就是完整方法论，直接能用。带 scripts/ 的技能（调研、规矩体检）把脚本拷到你的项目里，Python 直接跑（纯标准库，零依赖）。
Step by step / 分步说明：
- Pick a skill from the table below / 从下表找到解决你问题的技能
- Open its
SKILL.md— that's the full methodology / 打开它的SKILL.md——这就是完整方法论 - Hand it to your AI, and run the verification steps it prescribes — that's the whole point / 交给你的 AI，并执行它要求的验证步骤——这正是这套方法论的核心
| Skill（技能） | What it solves（解决什么） | How to use（怎么用） |
|---|---|---|
| Multi-Source Project Recovery （多源项目恢复） |
Rebuild a project's true state from scattered evidence — chat logs, git, artifacts. No guessing, every conclusion sourced. 从零散的历史证据（聊天记录/git/产物/知识库）重建项目的真实状态，不靠模型记忆补全，每条结论都有出处 |
Read SKILL.md, follow its 8 stages.读 SKILL.md，按里面的 8 个阶段走 |
| Chinese Industrial Research （中国工业行业调研） |
Verifiable research on Chinese industries: GB standards, vendor specs, EIA reports — every data point back-to-source, no fabrication. 可验证的中国行业调研：GB 标准、厂商规格、环评报告、上市公司文档，每个数据点都能回源，绝不编造 |
SKILL.md + 5 stdlib-only scripts in scripts/SKILL.md + scripts/ 里的 5 个脚本（纯 Python，无依赖） |
| AI Redteam Self-Test （AI 红队自测） |
Security checkups for your own AI stack before launch: prompt injection, jailbreaks, malicious skills. Test only what you own, in isolation. 上线前给自己的 AI 体系做安全体检：记忆库投毒、模型越狱、技能漏洞，只测自己的、隔离环境、测完补防护 |
Pick 2-3 tools from the selection table in SKILL.md按 SKILL.md 的工具选型表选 2-3 个工具开测 |
| AI Rule Enforcement （AI 规矩强制机制） |
Make AI follow rules by mechanism, not by asking: a zero-cost health-check script + a git pre-commit hook that blocks violations. 让 AI 守规矩不靠自觉、靠物理拦截：零成本体检脚本 + git 提交前自动检查，违规直接阻止提交 |
Run scripts/rule_health_check.py; install scripts/pre-commit运行 scripts/rule_health_check.py；安装 scripts/pre-commit 钩子 |
| Multi-Agent Relay Pipeline （多 AI 传帮带流水线） |
Pass-the-baton multi-agent pipeline: workers → optimizer distills lessons → next workers inherit → independent auditor verifies. Experience compounds. 多轮多 Agent 生产时让经验接力传承：Worker 干活 → 优化 Agent 提炼经验 → 下轮 Worker 继承 → 审计 Agent 验收 |
Build your pipeline per SKILL.md's architecture按 SKILL.md 的架构搭你的流水线 |
| Expert Prompt Assets （两专家提示词资产） |
Two copy-paste expert prompts: Design Expert (idea → blueprint) + Planning Expert (goal → task plan). Zero install, any AI platform. 两个复制即用的专家提示词：设计专家（想法→蓝图）+ 规划专家（目标→任务计划），不绑定任何 AI 平台 |
Copy the prompt from prompts/, paste into any AI chat复制 prompts/ 里的完整提示词，粘贴到任意 AI 新对话 |
| Shared KB Governance （知识库治理体系） |
Rules for a knowledge base shared by multiple AIs and one human: structure, naming, flow, sync, evidence discipline — machine-checked. 多个 AI 和一个人共享一个知识库时的规矩：目录分工、命名、流转、同步、证据纪律 |
Configure your KB per SKILL.md's contracts按 SKILL.md 的契约配置你的知识库 |
verifiable-agent-skills/
├── README.md
├── LICENSE（MIT）
└── skills/
├── multi-source-project-recovery/
│ ├── SKILL.md
│ └── restore_openai_export_attachments.py
├── chinese-industrial-research/
│ ├── SKILL.md
│ └── scripts/（audit_doc_scan / cn_std_search / openstd_std_list / s360_search / spec_table_extract）
├── ai-redteam-self-test/SKILL.md
├── ai-rule-enforcement/
│ ├── SKILL.md
│ └── scripts/（rule_health_check.py + pre-commit hook）
├── multi-agent-relay-pipeline/SKILL.md
├── expert-prompt-assets/
│ └── prompts/（design-expert-prompt.md + planning-expert-prompt.md + USAGE.md）
└── shared-kb-governance/SKILL.md
- Proven in production / 实战验证：a 30+ industry research pipeline, 4 rounds of full-coverage audits (5,000+ URLs, 20,000+ data points back-to-source), project recovery from fragmented chat archives, and a knowledge base governed by machine-enforced rules — every method comes from real incidents. 30+ 行业调研流水线、4 轮全量审计（5,000+ 个 URL、20,000+ 个数据点全部回源）、从碎片化聊天记录重建项目、机器强制规矩的知识库——每个方法都来自真实踩坑。
- No fabrication / 不编造：unknown is marked 【待核实】(to be verified) — never invented. 查不到就标【待核实】（待验证），绝不编数字。
- Evidence over assertion / 证据说话："done" requires a file path, a commit hash, or script output. If you can't write the evidence, you didn't do it. "做完了"必须能给出文件路径、提交号或脚本输出，给不出来就是没做。
This repo contains all 7 skills. If you only need one, check my profile for the standalone repos (names match the English names in the table above). 本仓库包含全部 7 个技能。如果你只需要其中一个，请到我的主页的仓库列表里找对应的独立仓库（名字与上表括号里的英文名一致）。
MIT. Any upstream tooling used by a skill is credited inside that skill's files. MIT。每个技能若使用了上游开源工具，均已在文件中注明来源。