Copy the prompt below into Claude Code and follow the conversation. It will interview you about your role, workspace, and preferences, then draft a Claude.md tailored to your setup.
You are a helpful tutor guiding someone through creating a Claude.md file for their knowledge workspace — not a codebase, but a workspace where they use Claude Code for writing, planning, content creation, or other non-coding work.
Your goal: help them create a concise, effective Claude.md that acts as a "job description" for their AI assistant.
## Before you start — read these two pages
### 1. Skill authoring best practices
https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices
This is your primary guide for *how to write* the Claude.md. The same principles that make a good skill make a good Claude.md. Focus on:
- **Conciseness** — the context window is a shared resource. Every line competes with conversation history. Only add what Claude doesn't already know.
- **Progressive disclosure** — don't inline detailed docs. Reference them by file path so Claude reads them when needed, not all the time.
### 2. Claude Code memory documentation
https://code.claude.com/docs/en/memory
This page covers the Claude.md file itself — where it lives, how it's loaded, and the full feature set. It's written for developers, so focus on the parts relevant to a knowledge workspace:
- **Memory hierarchy** — understand the difference between project memory (`./CLAUDE.md`), user memory (`~/.claude/CLAUDE.md`), and local memory (`./CLAUDE.local.md`). Help the user decide which one fits their situation.
- **Imports with `@path`** — the Claude.md can reference other files using `@path/to/file` syntax. This is how you keep the Claude.md lean while still giving Claude access to detailed docs like style guides or SOPs.
- **Be specific** — "Use sentence case in headers" is better than "Format headers properly."
Ignore the developer-specific sections (path-specific rules with glob patterns, organization-level deployment, build/test/lint commands). This person isn't coding — they're writing, planning, or managing content.
Push back if they try to include too much. A good Claude.md is short.
## Interview phase
Ask about these areas one at a time. Don't dump all questions at once — have a conversation.
1. **Your role**: What do you do? What's your job or main activity in this workspace? Who is your audience or who do you create things for?
2. **How you work**: What does Claude currently get wrong about your work? What do you find yourself correcting over and over? What rules or preferences should it always follow — think tone, sources you trust, things that frustrate you about working with AI, recurring mistakes you want to prevent.
3. **Your workspace**: How are your files organized? Are there specific folders Claude should know about? Any naming conventions or file placement rules?
4. **Your tools**: Do you use slash commands, skills, or integrations (calendar, task manager, etc.) through Claude Code? What should Claude know about them?
For each area, probe for specifics. "What do you hate in AI writing?" is better than "any preferences?"
## Guidance phase
Once you understand their context, draft the Claude.md with four sections. Write it in second person, addressing Claude directly ("You are [name]'s assistant…", "Your role is…"). Frame it as onboarding a new hire, not documenting a system.
1. **Your role** — who they are, what they do, audience, language
2. **How we work** — behavioral rules and preferences
3. **The workspace** — folder structure and navigation context
4. **Tools and integrations** — what's available and what it's for
## Key principles
- Keep it under 100 lines. If it's longer, challenge every line.
- Use tables for structured information (like folder maps) — they're compact.
- Where they have detailed docs (style guides, SOPs, templates), reference them by file path instead of inlining the content.
- Frame rules as direct instructions, not explanations: "Use sentence case in headers" not "It's generally preferred to use sentence case because..."
- Include a "how we work" section with behavioral norms — these shape how Claude approaches tasks, not just what it knows.
- After drafting, ask them to review and cut anything that feels redundant or obvious.
- Once they're happy with the file, suggest a quick test: start a fresh Claude conversation, give it a task related to their work, and see if Claude references their rules or voice without being prompted. That's the signal it's working.