What it is

A markdown file that lives in your project root. Claude reads it automatically at the start of every session. It is the single source of truth about you, your business, and how Claude should behave.

What to put in it

• Your business context — what you do, who you serve, how you price
• Your active client list and their status
• Your tech stack and tools
• How Claude should behave — be direct, no fluff, ask before committing
• What to avoid — no emojis, no over-promising, no vague suggestions

Structure

Three core sections that cover everything Claude needs:

• Who I am — identity, values, communication style
• Current projects — clients, products, priorities
• How Claude should work with me — rules, preferences, things to avoid

Example

# CLAUDE.md

## Who I Am
Name: Walid Boulanouar
Business: AY Automate — AI automation agency
Role: Founder, builder, systems thinker

## Current Projects
- 13 active clients across EU, US, MENA
- Building Humanoidz venture studio
- Scaling to 20 clients / 80K MRR

## How Claude Should Work With Me
- Be direct. No fluff. No apologies.
- Ask clarifying questions before assuming.
- Never commit to git without asking.

Time to set up

30 minutes for a solid first version. Iterate as your business evolves.

What it is

A file stored at ~/.claude/projects/.../memory/MEMORY.md. Claude reads it automatically every session and can write to it when you say "remember this." It grows with you.

Four memory types

• User memory — who you are, preferences, quirks that Claude learns
• Feedback memory — corrections you have made. "Never do X again." "Always format it like Y."
• Project memory — ongoing work context. Client statuses, sprint progress, blockers
• Reference memory — where to find things. File paths, tool configs, API endpoints

How Claude uses it

Read at session start alongside CLAUDE.md. When you tell Claude to remember something, it appends to MEMORY.md. Next session, that context is already loaded.

What makes it powerful

• No manual copy-pasting between sessions
• Corrections stick — tell Claude once, it remembers
• Project state carries forward — no re-explaining what happened last week
• Reference paths mean Claude finds files without asking you

What MCPs are

Model Context Protocol servers. They are bridges between Claude and external tools. Each MCP exposes specific actions — read, write, search, send — that Claude can call directly from the terminal.

The 8 MCPs

• Attio — CRM. Read contacts, deals, pipeline. Update records.
• Slack — Comms. Read channels, send messages, search history.
• Notion — Docs. Read pages, query databases, create entries.
• Unipile — LinkedIn. Send DMs, read profiles, publish posts.
• Anysite — Web. Fetch any URL, scrape content, read docs.
• GitHub — Code. Read repos, create PRs, manage issues.
• n8n — Automation. Trigger workflows, read executions, manage nodes.
• Filesystem — Files. Read, write, search across your local system.

How it works in practice

You say "check the pipeline in Attio for deals closing this week." Claude calls the Attio MCP, reads the data, and returns a summary. No browser, no tab-switching, no copy-pasting.

What hooks are

Shell scripts that trigger on specific Claude actions. They run before a tool is used (PreToolUse), after it completes (PostToolUse), or when the session ends (Stop). Think of them as guardrails.

The 6 hooks

• bash-safety.sh — Blocks destructive git commands: force push, reset --hard, branch -D, rm -rf. Prevents accidental data loss.
• file-protection.sh — Blocks edits to .env files and credentials. Warns before modifying CLAUDE.md, MEMORY.md, or settings.json.
• file-backup.sh — Creates a backup copy before every file edit. Keeps the last 10 versions in ~/.claude/backups/.
• search-year.sh — Automatically appends the current year to web searches that are missing a date. Keeps results fresh.
• file-logger.sh — Logs every file change to a daily changelog. Full audit trail of what Claude touched.
• stop-notify.sh — Plays a sound when Claude finishes working. Walk away and come back when it is done.

How they work together

Before Claude runs git push --force, bash-safety blocks it. Before Claude edits a config file, file-protection warns you. After every edit, file-backup saves a copy. The system protects itself.

What skills are

Markdown files stored in .claude/skills/ that define a complete Claude workflow. Each skill has a trigger phrase, input requirements, step-by-step instructions, and output format. One command triggers the full workflow.

6 categories

• Sales — /personalized-value-outreach — Research a prospect, score against ICP, draft a value-first DM, send via Unipile
• Content — /yt-to-content — Extract a YouTube transcript, generate 3 LinkedIn posts in different formats
• Client — /meeting-processor — Process a meeting transcript into action items, client notes, and follow-up tasks
• Intelligence — /content-intelligence — Analyze your LinkedIn performance, extract patterns, suggest what to double down on
• Technical — /excalidraw-diagram — Generate architecture diagrams from text descriptions
• Operations — /sop — Turn any task into a step-by-step Standard Operating Procedure for the team

How skills use the other layers

A skill reads CLAUDE.md for your voice and business context. It checks MEMORY.md for client history. It calls MCPs to fetch data from Attio or send messages via Unipile. Hooks keep it safe the entire time.