How real projects brief Claude Code: the commands, conventions and warnings they put in CLAUDE.md.
AnthropicCLAUDE.mdanthropics/claude-cookbooks/claude_agent_sdk/chief_of_staff_agent/CLAUDE.md
Output: JSON with monthlyburn, runwaymonths, totalrunwaydollars, quarterlyburn, and burnrate_daily Remember: As Chief of Staff, you have access to financial data in the financial_data/ directory and can delegate specialized analysis to your subagents (financial-analyst and recruiter).
AnthropicCLAUDE.mdanthropics/financial-services/CLAUDE.md
Cowork plugins and Claude Managed Agent templates for financial services. Each named agent ships two ways from one source. Run python3 scripts/check.py before committing — it lints every manifest, verifies all system.file / skills.path / callable_agents.manifest references resolve, fails if any agent-plugins/<slug>/skills/ copy has drifted from its vertical-plugins/ source, and rejects non-ASCII bytes in a .ps1 without a UTF-8 BOM. Keep .ps1 files pure ASCII. Windows PowerShell 5.1 — still the default shell on managed Windows — decodes a BOM-less…
AnthropicCLAUDE.mdanthropics/knowledge-work-plugins/partner-built/slack/CLAUDE.md
This plugin integrates Slack with Claude Code, providing tools to search, read, and send messages in Slack.
AnthropicCLAUDE.mdanthropics/claude-quickstarts/managed-agents/chat-sdk/CLAUDE.md
Long-running Node server (Hono): browser chat with a sessions sidebar (useChat → /api/chat) → Chat SDK web adapter onDirectMessage → one Managed Agents session per conversation, where the conversation ID is the session ID (no mapping, no database; the sidebar is sessions.list, transcript replay is sessions.events.list) → a held session event stream (opened with event_deltas token previews, always on) feeds each fragment back onto the same HTTP response while /api/activity streams the tool calls and thinking the page shows as…
AnthropicCLAUDE.mdanthropics/claude-quickstarts/managed-agents/knowledge-wiki/CLAUDE.md
This directory is a Managed Agents cookbook: it builds a real M&A data room into a knowledge wiki (a versioned memory store) and queries it. The runnable artifact is distilldocumentsintoknowledgewiki.ipynb; everything else exists to set that notebook up. When a user asks you to walk them through setting this quickstart up, read skill.md in this directory and follow it step by step. Do not improvise the setup — the SDK build, the SEC user-agent, and the fetch order all matter.
AnthropicCLAUDE.mdanthropics/claude-quickstarts/managed-agents/mcp-server-typescript/CLAUDE.md
MCP server that exposes Managed Agents session primitives as tools. Two entrypoints sharing one tool set: src/server.ts (stdio → Claude Desktop and Claude Code) and src/server-http.ts (Streamable HTTP → claude.ai custom connector). This quickstart creates no agents, so there is no agents/ directory. It needs one environment, created once from environment.yaml (the command is in that file and in README.md). Commands: bun run stdio (Desktop, Code), bun run http (claude.ai), bun run typecheck.
AnthropicCLAUDE.mdanthropics/claude-quickstarts/managed-agents/roadtrip-planner/CLAUDE.md
A Next.js chat built directly on a Managed Agent session: no chat framework, the event log is the app state. Teaches session event streaming (eventdeltas), vault credential injectionlocation, per-session agentwithoverrides, and multiagent coordinator rosters. The planner calls the National Park Service and Windy APIs with keys it never holds, then hands its draft to a second agent for a quick critique. That second agent is an Opus reviewer running as a session thread. 1. Invoke /claude-api first. It loads the…
AnthropicCLAUDE.mdanthropics/claude-quickstarts/managed-agents/sentry/CLAUDE.md
Scheduled deployment: cron → Managed Agents session with sentry-cli and a vault env-var credential → triage report in /mnt/session/outputs/. Pull exact shapes from the /claude-api skill's shared/managed-agents-*.md docs. Provisioning is ./agents/setup.sh (ant 1.34 or later, jq): it runs ant apply --yes agents environments vaults, which creates the agent, environment, and vault from those directories and records their IDs in claude-lock.json, then adds the Sentry token to the vault as an environmentvariable credential (skipped when the vault already has one), and…
A file Claude Code reads at the start of every session. It holds the commands, conventions and warnings the agent needs for this project.
At the repository root. Claude Code also reads CLAUDE.md files in subdirectories when it works there.
Build and test commands, the project's layout, conventions that aren't obvious from the code, and mistakes to avoid. Short files tend to work better than long ones.
Claude Code reads CLAUDE.md; most other agents read AGENTS.md. Many projects keep one and point the other at it.
