AGENT SKILL

reverification

By-hand only — DO NOT auto-trigger. When the user types /reverification (or asks for "reverify this", "independent verification", "LLM-uncheatable check", "第三方再验证", "用 repo 之外的方法验证"), re-verify recently-completed work using verification methods that **existed before this repo existed** and that an LLM agent cannot fabricate. Pick 3-5 verifications from a 10-family menu (hardware / cryptographic / blockchain / external real-system / mathematical / temporal / content authenticity / physical / network / third-party trust), dispatch them via real tools that capture raw evidence, and report PASS / FAIL / INCONCLUSIVE based ONLY on captured evidence — never on internal reasoning. Refuse to fall back on TeamBrain's bespoke verifiers (judge harness, fastprobe, /review, feature-verification gates) because those did NOT exist before the repo.

License 许可
MIT
最近更新
2026-05-20
Stars
★ 1
Forks
0

01 前言 Front Matter

name
reverification
triggers
[ "/reverification", "reverify this", "reverify the work", "independent verification", "LLM-uncheatable check", "pre-existing verification", "第三方再验证", "用 repo 之外的方法验证", "不让 LLM 作弊的验证" ]
description
By-hand only — DO NOT auto-trigger. When the user types /reverification (or asks for "reverify this", "independent verification", "LLM-uncheatable check", "第三方再验证", "用 repo 之外的方法验证"), re-verify recently-completed work using verification methods that **existed before this repo existed** and that an LLM agent cannot fabricate. Pick 3-5 verifications from a 10-family menu (hardware / cryptographic / blockchain / external real-system / mathematical / temporal / content authenticity / physical / network / third-party trust), dispatch them via real tools that capture raw evidence, and report PASS / FAIL / INCONCLUSIVE based ONLY on captured evidence — never on internal reasoning. Refuse to fall back on TeamBrain's bespoke verifiers (judge harness, fastprobe, /review, feature-verification gates) because those did NOT exist before the repo.
allowed-tools
[ "Bash", "Read", "WebFetch", "WebSearch", "Grep", "Glob", "Agent" ]

03 SKILL.md 正文

/reverification — 第三方再验证(pre-existing & LLM-uncheatable)

   __                                                          .---.
 _(o o)<   呷呷~ 鸭鸭是 by-hand only 的再验证向导            ( raw  )
  \   )     不写死步骤、不调 repo 自家工具                   ( evi  )
   `-'      由 LLM 现场从「repo 还没出生时就存在」的          ( dence )
            验证宇宙里挑 3-5 个,跑出 raw evidence,          `-----'
            再用 evidence 反推 PASS / FAIL                       ↑
                                                                 │
                                                          外部物理 / 密码学
                                                          / 网络 / 第三方

灵魂条款(不要违反)

这只 skill 不写死任何脚本、任何固定命令、任何固定顺序。调用时的 LLM 自己负责挑、自己负责跑、自己负责诚实承认抓不到证据时的失败。

为什么这样?因为 TeamBrain 自家的验证体系(judge harness / fastprobe / /review / feature-verification gates)都是 repo 出生之后才有的,LLM 完全可以理解它们的内部逻辑并伪造看起来合理的 output。「pre-existing & uncheatable」的本质是:验证结果依赖 LLM 触碰不到的外部现实 —— 它不知道某个信息,而这个信息必须通过物理世界、密码学机制、外部网络或第三方信任才能获得。

优先级链:硬件锚定 > 密码学承诺 > 外部真实系统 > 数学决定性证明。

验证宇宙(10 家族菜单 —— LLM 现场选,不是清单)

# 家族 代表手段 为什么 LLM 作弊不了
1 Cryptographic / hash / signature Merkle 证明、RSA/ECDSA、PKI 链、RFC 3161 时间戳、HMAC、ZKP、homomorphic hash 私钥 / 共享密钥 / 真实数据 hash 链 LLM 拿不到
2 Hardware-backed security TPM EK、SGX/TrustZone enclave、HSM、Secure Boot、TEE remote attestation 芯片级信任根,飞地内存不可读
3 Blockchain / distributed consensus PoW 链历史、智能合约状态、ENS、IPFS CID、Sigstore/Rekor 签名 commit 共识 / 透明日志需多方算力,LLM 单方面改不动
4 External real-system APIs 天气、ADS-B 航班、交易所 order book、卫星图像、IoT 传感器、医疗 telemetry、SCADA 真实物理信号 / 实时下行数据,LLM 接不到
5 Mathematical decidable proofs SAT/SMT solver、Coq/Lean、DH/ECDH、KZG commitment、SNARK 机器证明 + kernel 校验,LLM 绕不过
6 Temporal / ordering NTP、Lamport 向量、FIFO 队列、VDF(verifiable delay function)、可验证延迟编码 真实时间流逝,LLM 在推理时压缩不了
7 Content / media authenticity EXIF、相机 RAW、视频关键帧 hash、音频声纹、FASTQ DNA reads、质谱原始数据 仪器硬件签名 / 物理传感器原始噪声分布
8 Physical / geographic GPS+原子钟、RFID/NFC PUF 标签、邮政追踪、SWIFT 电文、公证档案、法院案号 卫星定位 / 物理不可克隆函数 / 政府登记
9 Network / transport TCP seq number、TLS handshake 机密、DNSSEC、mTLS 双向证书 三次握手随机数 / pre-master secret LLM 预测不到
10 Third-party trust anchors Sigstore/Rekor 透明日志、CWT、OCSP 吊销、W3C VC、Slack/Discord server 时间戳 多个独立信任源同时伪造的成本

完整 55+ 条目录参见用户 claudefast -p "list ... LLM cheat ..." 的输出(每次调用重新生成,不要在这里硬列)。

调用流程(不是脚本,是 playbook)

  1. 读 context — 先看最近改了什么:git diff main...HEAD / 最近 Write/Edit 历史 / 当前要验证的 artifact 是文件、URL、binary、commit、还是 release tag。
  2. 挑 3-5 个验证 —— 从 10 家族里挑。每挑一个,写两行 justification:
    • (a) 为什么这个验证对这个 artifact 相关(不是泛泛而谈)。
    • (b) 为什么 LLM 自己没法伪造 verification 的 output(如果想不出 (b),换一个)。
  3. 跑 + 抓 raw evidence —— 用 Bash 跑真实命令、WebFetch 打真实第三方 URL、Agent 派 explore subagent 拿独立 diff 阅读结果、或 mcp__gbrain__* 查外部 brain。raw evidence 写到 .reverification/<timestamp>/<verification-id>/ 下,包括 stdout、response body、签名 blob、JSON、截图路径。
  4. read-only 综合 —— 只在 evidence 落盘之后写 PASS / FAIL / INCONCLUSIVE 报告。报告里每个结论必须引用具体 evidence 文件路径,禁止「我觉得 / 看起来 / 应该」式判断。
  5. 诚实输出 —— 如果 3-5 个验证里没有任何一个抓到 cheat-proof evidence,结论必须是 FAIL 或 INCONCLUSIVE。不要用「大体看下来 OK」之类语言搭桥。

反模式(必须拒绝)

  • ❌ 写 scripts/reverify-*.sh —— 违反 user memory feedback_judge_harness_md_playbook.md。这只 skill 永远是 MD playbook,不是 bash 脚本生成器。
  • ❌ 调 TeamBrain 自家验证(judge harness / fastprobe / /review / feature-verification gates / claudefast snapshot 等)—— 它们 repo 出生之后才有,不符合「pre-existing」前提。
  • ❌ 没 evidence 就下结论 —— 哪怕只 1/5 抓到 evidence 也得在报告里明确说「其余 4 个未能抓到 cheat-proof evidence」并降级为 INCONCLUSIVE。
  • ❌ auto-trigger —— 必须用户显式喊 /reverification 或同义短语才动;SessionStart / Stop / PostToolUse 之类自动钩子里出现这只 skill 视为 bug。
  • ❌ pnpm typecheck / pnpm test / pnpm build 单独当答案 —— 这些工具确实 pre-existing,但 test 文件本身可以被作者写成「永远绿」的样式,LLM 在写代码时一并改测试就能骗过。所以即便选了它们,也必须组合至少一个更强的锚点(签名 / 网络 / 时间戳 / 第三方 server)。

输出格式(建议,非强制)

# /reverification 报告 — <artifact 描述> — <UTC 时间戳>

## 选了哪几个验证(justified)
1. <家族 #X · 手段名> — relevance: <一行>  /  uncheatability: <一行>
2. ...

## raw evidence 落盘位置
- .reverification/<ts>/01-<id>/{stdout.txt, response.json, ...}
- ...

## 逐项判定
- 验证 1:PASS / FAIL — 引用 `<evidence path>:line`
- ...

## 综合结论
PASS / FAIL / INCONCLUSIVE — 一句话,引用上面的逐项判定。

呷呷~ 鸭鸭说完了,由你(运行时 LLM)现场挑、现场抓证据、现场承担诚实义务。(>ω<)

04 同仓库更多 Skills

libz-renlab-ai
libz-renlab-ai/Matrix-Viki

zoom-out

Tell the agent to zoom out and give broader context or a higher-level perspective. Use when you're unfamiliar with a section of code or need to understand how it fits into the bigger picture.

★ 1
libz-renlab-ai
libz-renlab-ai/Matrix-Viki

triage

Triage issues through a state machine driven by triage roles. Use when user wants to create an issue, triage issues, review incoming bugs or feature requests, prepare issues for an AFK agent, or manage issue workflow.

★ 1
libz-renlab-ai
libz-renlab-ai/Matrix-Viki

to-prd

Turn the current conversation context into a PRD and publish it to the project issue tracker. Use when user wants to create a PRD from the current context.

★ 1
libz-renlab-ai
libz-renlab-ai/Matrix-Viki

to-issues

Break a plan, spec, or PRD into independently-grabbable issues on the project issue tracker using tracer-bullet vertical slices. Use when user wants to convert a plan into issues, create implementation tickets, or break down work into issues.

★ 1
libz-renlab-ai
libz-renlab-ai/Matrix-Viki

reverification

By-hand only — DO NOT auto-trigger. When the user types /reverification (or asks for "reverify this", "independent verification", "LLM-uncheatable check", "第三方再验证", "用 repo 之外的方法验证"), re-verify recently-completed work using verification methods that **existed before this repo existed** and that an LLM agent cannot fabricate. Pick 3-5 verifications from a 10-family menu (hardware / cryptographic / blockchain / external real-system / mathematical / temporal / content authenticity / physical / network / third-party trust), dispatch them via real tools that capture raw evidence, and report PASS / FAIL / INCONCLUSIVE based ONLY on captured evidence — never on internal reasoning. Refuse to fall back on TeamBrain's bespoke verifiers (judge harness, fastprobe, /review, feature-verification gates) because those did NOT exist before the repo.

★ 1
libz-renlab-ai
libz-renlab-ai/Matrix-Viki

repo-issues-status

TeamBrain 仓库**全量 issue 全景**报告(all-scope only)。仅当用户问"status of all issues" / "show me all issues" / "what issues are open" / "all open issues" / "issue 状态(全景)" / "所有 issue 状态" / "issue 全景" / "仓库 issue 概览" / "还有哪些 open issue" 这类**仓库级、多 issue、概览类**问题时触发。**不要在以下情况触发**:(a) 用户问单个 issue 的状态("what's the status of issue

★ 1