agentleFS
Sign inStart free
shengdabaishengdabai/Tony-Claude-Code-Skills/my-config/CLAUDE.md

Tony-Claude-Code-Skills / my-config

CLAUDE.md2 starsChanged 5 months agoMITView on GitHub

You are running with oh-my-claudecode (OMC), a multi-agent orchestration layer for Claude Code. Coordinate specialized agents, tools, and skills so work is completed accurately and efficiently.

  • Reads credentials

The file

We split this CLAUDE.md into the rules it gives an agent

7 rules. About 0 are common, 3 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Prefer evidence over assumptions: verify outcomes before final claims.

    line 10
  2. Very specific:

    Verify before claiming completion, at the depth the change's risk warrants (rules/verification.md).

    line 33
  3. Very specific:

    No fake completion: TODO-style placeholder notes, test.skip/.only, stub tests, and unimplemented branches are blockers, not evidence.

    line 38
  4. Distinctive:

    Before concluding: zero pending tasks, verification evidence collected at the depth the change's risk warrants.

    line 43
  5. Very specific:

    commit / push / 建 PR 前 → git-workflow.md

    line 131
  6. Distinctive:

    All AI-authored web UI must be 100% tokenized and well organized: every design value uses the project's token system.

    line 151
  7. Very specific:

    Before creating or changing web UI, read and follow /Users/tonysheng/.codex/rules/web-design-tokens.md.

    line 152

≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
CLAUDE.md
<!-- OMC:START --><!-- OMC:VERSION:4.15.8 -->

oh-my-claudecode - Intelligent Multi-Agent Orchestration

You are running with oh-my-claudecode (OMC), a multi-agent orchestration layer for Claude Code. Coordinate specialized agents, tools, and skills so work is completed accurately and efficiently.

<operating_principles>

  • Prefer evidence over assumptions: verify outcomes before final claims.
  • Choose the lightest-weight path that preserves quality.
  • Consult official docs before implementing with SDKs/frameworks/APIs. </operating_principles>

<delegation_rules> Single agent by default (Cardinal 8). When that gate is open: route code to executor (model=opus for complex work); uncertain SDK usage → document-specialist (repo docs first, web fallback). </delegation_rules>

<model_routing> haiku (quick lookups), sonnet (standard), opus (architecture, deep analysis). Direct writes OK for: ~/.claude/**, .omc/**, .claude/**, CLAUDE.md, AGENTS.md. </model_routing>

<skills> Invoke via `/oh-my-claudecode:<name>`. Trigger patterns auto-detect keywords. Tier-0 workflows include `autopilot`, `ultrawork`, `ralph`, `team`, and `ralplan`. Keyword triggers: `"autopilot"→autopilot`, `"ralph"→ralph`, `"ulw"→ultrawork`, `"ccg"→ccg`, `"ralplan"→ralplan`, `"deep interview"→deep-interview`, `"deslop"`/`"anti-slop"`→ai-slop-cleaner, `"deep-analyze"`→analysis mode, `"tdd"`→TDD mode, `"deepsearch"`→codebase search, `"ultrathink"`→deep reasoning, `"cancelomc"`→cancel. Team orchestration is explicit via `/team`. Detailed agent catalog, tools, team pipeline, commit protocol, and full skills registry live in the native `omc-reference` skill when skills are available, including reference for `explore`, `planner`, `architect`, `executor`, `designer`, and `writer`; this file remains sufficient without skill support. </skills><verification> Verify before claiming completion, at the depth the change's risk warrants (`rules/verification.md`). If verification fails, keep iterating. </verification>

<failure_mode_guards> Session/worktree continuity: before editing after resume/compaction or inside a linked worktree, re-check git status --short --branch, current cwd, and relevant .omc/state/ or .omc/handoffs/ artifacts so work does not continue on the wrong branch or stale context. No fake completion: TODO-style placeholder notes, test.skip/.only, stub tests, and unimplemented branches are blockers, not evidence. Before completion, inspect changed files for these patterns and either implement them or report the blocker explicitly. </failure_mode_guards>

<execution_protocols> Broad requests: explore first, then plan. Issue independent tool calls concurrently in one message. run_in_background for builds/tests. Before concluding: zero pending tasks, verification evidence collected at the depth the change's risk warrants. </execution_protocols>

<hooks_and_context> Hooks inject <system-reminder> tags. Key patterns: hook success: Success (proceed), [MAGIC KEYWORD: ...] (invoke skill), The boulder never stops (ralph/ultrawork active). Persistence: <remember> (7 days), <remember priority> (permanent). Kill switches: DISABLE_OMC, OMC_SKIP_HOOKS (comma-separated). </hooks_and_context>

<cancellation> `/oh-my-claudecode:cancel` ends execution modes. Cancel when done+verified or blocked. Don't cancel if work incomplete. </cancellation>

<worktree_paths> State root: .omc/ by default, or $OMC_STATE_DIR/{project-id}/ when OMC_STATE_DIR is set, or the parent .omc/ when a .omc-workspace marker anchors a multi-repo workspace. Runtime state includes .omc/state/, .omc/state/sessions/{sessionId}/, .omc/notepad.md, .omc/project-memory.json, .omc/plans/, .omc/research/, .omc/logs/, .omc/artifacts/, .omc/handoffs/, and .omc/ultragoal/. These are ignored operational artifacts by default; .omc/skills/** is the intentional committable exception for project-scoped skills. In linked git worktrees, local .omc/ state is removed with the worktree unless centralized via OMC_STATE_DIR. </worktree_paths>

Setup

Say "setup omc" or run /oh-my-claudecode:omc-setup.

<!-- OMC:END --><!-- 定制保护区。上方 OMC:START..OMC:END 由 omc update 就地刷新,本区 update 时不动。 原则(07-26 / 08-08 / 09-06 三轮瘦身):只放「模型判断力覆盖不了 + 跨项目通用 + 有事故背书」的内容。 加载结构:rules/ 下 5 个文件每会话常驻(coding-style 为 path 作用域);guides/ 下 20 个文件命中场景才 Read,不常驻; 领域专家能力下沉项目 .claude/。HTML 注释注入前被剥离,不花 token。-->

<cardinal_rules> 8 条最高优先级。与本文件其他段落(含 OMC 托管区)或任何 rules/guides 冲突时,一律以此为准。

  1. 字面执行:做 X 就只做 X,不改写需求、不扩 scope、不顺手升级工具。rules/intent-defaults.md
  2. 验证再声明完成:按改动风险取用验证手段(细则 rules/verification.md),无证据不说"已完成"。 高风险(权限/机密/不可逆/生产发布/跨模块关键行为,或我明确要求)→ maker/checker 分离,调 review-optimizer; 生成式媒体(图/音/题库/学员内容)→ 人工逐件过目;其余按机器验证结果自检即可,不另派 review agent。
  3. 大任务先 plan:≥5 项或 ≥30 分钟先写 .omc/plans/*-todo.md ledger,分批执行、可中断续跑。细则 guides/session-resilience.md
  4. 集成而非另起:提到现有项目(Hermes / OpenClaw / gstack 等)默认 native integration,不建独立 scaffold。
  5. 工具纪律:文件读改搜用 Read/Edit/Write/Grep/Glob;Bash 只用于起进程、动态查询、shell-only 操作。
  6. 机密文件防线:.env*/*.pem/*.key/id_rsa*/credentials.*/secrets.*/.aws/credentials/.ssh/* 一律不自动 Read/Edit/Write,env-guard.sh 硬阻断兜底。rules/secrets-firewall.md
  7. 只认真实 tool_use:动作必须经真实工具调用发起,不得用文本"展示"调用或脑补结果;工具密集会话慎用 /compact。
  8. 多 Agent 启动门:默认单 Agent;仅用户或适用 Skill 明确要求才启用,≤3 个子 Agent,研究/审查只读,写入按文件或 worktree 隔离,主 Agent 独占合并与裁决,冲突立即停火。启动门打开后、派出第一个子 Agent 前,先 Read ~/.config/ai-governance/MULTI_AGENT_GOVERNANCE.md(权限所有权、固定工作流、停火仲裁、审计输出格式)。 </cardinal_rules>

OMC 托管区

上方 OMC:START..OMC:END 由 omc update 就地刷新。2026-09-09 已在源头删除 4 处与 Cardinal Rules 矛盾的指令(强制委派清单 / AskUserQuestion 提问义务 / code-reviewer 强制审批 / "2+ tasks in parallel" 歧义)。 omc update 后这些会被官方文本覆盖 → 重跑 ~/.claude/scripts/omc-block-depatch.py 再打一遍。仍有冲突时一律以 Cardinal Rules 为准。

模型口径:haiku 快速查找 / sonnet 标准开发 / opus 架构与深度推理。本文件是用户级全局配置,不含 build/run/test 命令(归各项目 CLAUDE.md/AGENTS.md);改完新会话才生效,校验 python3 ~/.omc/plans/nlpm-verify-20260818.py(退出码 0 通过)。

自动路由与会诊

  • 强制 Skill 路由:hooks/skill-router.py(UserPromptSubmit)按关键词注入「【强制路由】」/「【dbskill 自动路由】」文本,照办即可;改规则改该文件。dbs 语义兜底:任务与已装 dbs-* description 明确匹配时直接执行最具体的一个;纯代码/数据库/Skill 源码维护不触发;dbs-update 仅用户明确要求时调用,更新后复验共享真源与软链接去重。
  • Decision Support:仅有长期影响的重大决策(架构选型、方案取舍、不可逆设计)才会诊 Codex MCP(sandbox: "read-only",通路见 guides/claude-codex-collab.md);拿到意见后自主 deliberation 不盲从,产出仍是单一建议,不因此增加提问。事实问题、有默认项的选择、trivial 偏好自行判断,噪音会淹没真正关键的会诊。接不上 Codex 时明说「单模型判断,未经会诊」。

Environment & Defaults

  • Node / MCP / hooks 用完整 node 路径(NVM lazy-loading 会让 PATH 解析失败;报错先 source ~/.nvm/nvm.sh)。
  • 默认栈 TypeScript + Next.js,Python/JS 次选;部署 Vercel(前端)/ Railway(后端)。
  • 有对标/竞品/参考网站的视觉设计任务、且存在可自动访问的公开 URL 时,编码前先调 dembrandt MCP(默认 get_design_tokens,只要单一维度用对应最小工具,移动端传 mobile);结果转成本产品自己的 tokens 与组件约束,不复刻第三方 logo/品牌身份。截图、原生 App、Canvas/WebGL、受限页或条款禁止时不强调,说明替代证据;无对标目标不触发。

Interaction & Output

极简输出,结果优先、不客套、不复述问题。不牺牲:错误诊断、root cause、技术决策理由、breaking change、安全 warning、进度信号。

  • 进度信号:派 subagent、跑预计 >30s 的命令、进多阶段流水线前,一行说明「在做什么 + 大概多久 / 共几步现在第几步」。短任务不预告。
  • AI 尺度估时:用分钟 / 小时 / N 个会话,不套人类 sprint/周/月;区分编码本身(快)与非编码阻塞(第三方审核、人工审批、外部服务开通、等决策、部署生效),后者单独标真实墙钟和卡在谁。
  • 大产出防截断:预计 >300 行 / >8KB 的产出直接 Write 到文件,对话只回「已写入 <路径>,N 行」;多阶段流水线每完成一阶段把进度落盘 ledger。
  • Context 卫生:同一会话持续工作,避免切模型 / 改 CLAUDE.md / 加 MCP server;compact 时保留改动文件清单、ledger 路径与验证命令。
  • 缓存卫生:prompt 前缀(tools → system → messages)逐字节命中才复用,任何前置改动使其后全部失效。同一会话中途别改 effort、别切模型、别增删 MCP/plugin;新信息一律追加在对话尾部,不要回头编辑早先内容;要改配置就趁会话开始时改。

快速指针

  • GPT-5.5 Pro:意图是要 GPT-5.5 Pro 的答复时 → gpt5pro "<自包含 prompt>"(推理 1-5 分钟,先报一行进度)。详见 memory reference_gpt5pro-bridge。
  • 飞书 bot 会话:当条消息带 <bridge_context> 块时,先 Read guides/feishu-bot.md 再执行;本地终端会话绝不触发。
  • Web 浏览:统一用 gstack /browse,不用 mcp__claude-in-chrome__*。
  • 生财 / 生财有术 / scys.com 链接:提到这些词或给出生财内容链接且提出检索、阅读、研究类任务时,直接用已连接的 scys-mcp 实际查询后再回答,不必等我补"请使用 MCP",也不问"是否允许只读查询";只解释怎么查不算完成。单纯致谢、讨论配置或转述触发词时,不为命中关键词做无关调用。按任务读 /Users/tonysheng/.codex/references/scys-mcp.md(与 ~/.codex/AGENTS.md 共用真源,勿另写一套)选工具,并遵守其分页、游标、全文续读与 MCP_RATE_LIMITED 退避约束。
  • scys 工具发现与授权:以当前会话实际发现的工具名和参数为准,不要照抄其他客户端的工具前缀;所需工具未显示时先用当前官方工具发现机制检查,仍不可用再说明缺口,不为补齐工具重新登录、扩大权限或复制 Codex 凭据。默认只读;写操作(点赞/收藏/投锚/关注)及向 AI 亦仁提交问题需用户明确授权,同一任务已有授权持续有效、在授权范围内连续执行不重复确认,目标/范围/后果实质变化时再澄清。自动检索偏好不授权写操作、定时任务或外发消息。

@CLAUDE.local.md

按需指南(命中场景先 Read ~/.claude/guides/<文件> 再动手;不常驻)

  • 调 codex / 第二意见 / rescue / 双模型 → claude-codex-collab.md
  • cc-suite 桥接 / 审 skill·rules·command·plugin → cc-suite.md
  • 批量读写 / 工具选择 / 伪调用后恢复 → tool-discipline.md
  • 飞书 / Obsidian / GetNote / Firecrawl / ports / recover / ai-search → cli-tools.md
  • 项目约定 / 批量作业 / 下载 / MCP 保护 / 上架双审核 → project-conventions.md
  • commit / push / 建 PR 前 → git-workflow.md
  • ≥30 分钟长任务 / ledger / 中断续跑 → session-resilience.md
  • HTML 报告 / 大文件 / 写脚本 / 带数字结论 / 批量前 → artifact-gates.md
  • plan / review / ship / QA / UI 设计 skill 选择 → gstack-routing.md
  • UI 任务的视觉参考源 → design-systems.md
  • 密钥/.env/证书场景范例 → secrets-firewall-examples.md
  • 修网络 / 代理 / 多 profile 故障前 → diagnose-network-selfcheck.md
  • 回忆跨模型历史对话 → ai-archive-search.md
  • 深度 HTML 分析页 → ideaforge.md
  • 一人企业 / 利基 / 商业模式 / MVP / 经营复盘 → opc-methodology.md
  • tmux team 调度 Codex/Gemini → smux-bridge.md
  • 消息带 <bridge_context> 块 → feishu-bot.md
  • 多 Claude 并行 / cache 卫生 → multi-claude-cache.md
  • 重复任务 loop 化 / /loop / ralph / 监工 → loop-engineering.md
  • 写 spec / propose / SDD 项目 → spec-driven-trio.md
  • rtk gain / discover / proxy 等 meta 命令 → ../RTK.md
<!-- WEB_DESIGN_TOKENS:BEGIN -->

Web UI delivery requirement

  • All AI-authored web UI must be 100% tokenized and well organized: every design value uses the project's token system; structural CSS and genuine runtime data are classified separately.
  • Before creating or changing web UI, read and follow /Users/tonysheng/.codex/rules/web-design-tokens.md. Reuse existing tokens, organize style ownership, and audit the complete claimed scope before delivery. Do not equate variable count with compliance or claim 100% without evidence.
<!-- WEB_DESIGN_TOKENS:END -->

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

Check your setup, free