By the end of this chapter you can write a CLAUDE.md that sets the standing rules for a project, seed it from your personal constitution, and even have Claude draft it with you.
You know an .md file is just a note. Now meet the most useful one you will ever write. CLAUDE.md is a file with a special job: when it sits in a project, Claude reads it first, every time, and treats it as the standing law of that project. It is the project's constitution, the way Chapter 3 gave you a personal one.
This is where a real shift happens. You are not prompting better any more. You are building the world the AI works inside. A good prompt steers one answer. A good CLAUDE.md steers every answer, forever, without you saying it again. That is a different kind of power, and it is what building the world actually means.
AVATAR OPENER · ~90s
Watch: writing the file Claude reads before every single task
HeyGen avatar · generated, consistent presenter
Your personal constitution is about you and travels everywhere. CLAUDE.md is about one project and stays with it. They are the same idea at two scales. In fact the fastest way to write your first CLAUDE.md is to start from your personal constitution and add the project-specific rules on top. You are not starting from a blank page. You are inheriting from a document you already wrote.
The name CLAUDE.md is Claude's, but the idea is not. The same move exists elsewhere: ChatGPT Projects and Gemini Gems both let you set standing instructions a workspace always follows. Different name, different menu, same principle: write the rules once, let them shape every answer. Learn it here and you can carry it into any of them.
Prompting better
You craft one clever request. It steers exactly one answer. The next chat, you start over.
Result: good answers, one at a time, paid for each time.
VS
Building the world
You write the standing rules once in CLAUDE.md. They shape every answer in the project automatically.
Result: consistent behaviour, set once, free forever after.
You are not prompting better. You are building the world the AI lives in.
A prompt steers one answer. A CLAUDE.md steers every answer in the project, without you ever saying it again.
WHAT GOES IN IT
A first CLAUDE.md has four parts, and you have already written the raw material for most of them. Think of it as a briefing the project never forgets.
Who and what
What this project is and who you are, pulled from your constitution.
How to behave
The standing rules: tone, format, what to always do.
What never to do
The hard limits. The moves that would never be you.
When unsure
Tell it to ask one question before guessing.
SEE IT
Here is a starter CLAUDE.md you can copy today. It is built for a freelance writer, but the bones fit any project. Notice how it opens by pointing at the personal constitution: the project inherits you, then adds its own rules.
Starter CLAUDE.md (copy this)
# CLAUDE.md
## Who I am
Read my constitution file first and treat it as how I sound in everything.
Short version: I am a freelance writer for B2B software companies. Solo.
## What this project is
A workspace for drafting client articles and my own newsletter.
Default to the client unless I say it is for the newsletter.
## How to behave
- Match my voice. Plain, concrete, no buzzwords.
- Show your reasoning briefly before a long piece, so I can redirect early.
- When you cite a fact, say where it came from.
## Never
- Never invent a statistic, a quote, or a source.
- Never use the words leverage, synergy, or unlock.
- Never write a conclusion that just repeats the intro.
## When unsure
Ask me one clear question before writing. Do not guess at the brief.
And the best part: you do not have to write it cold. You can ask Claude to draft it with you. Tell it about your project, paste your constitution, and ask it to propose a CLAUDE.md. Then you edit. The tool writes the first draft of the rules it will follow, and you make the calls. That is directing, applied to the file that does the directing.
NOW YOU TRY · CREATE
Write your project's CLAUDE.md
Write a CLAUDE.md for your thread project. Start from your personal constitution (point at it in the first section), then add the four parts: what this project is, how to behave, what never to do, and the ask-first rule. If you get stuck, paste your constitution into a chat and ask Claude to draft a CLAUDE.md with you, then edit its draft until the rules are actually yours.
Right if someone could read your CLAUDE.md and predict how your project will behave on a task you have not given it yet.
Show the worked solution
The test of a CLAUDE.md is prediction. A good one lets a stranger guess how the project will act on a brand new task. Take the writer's file above. Hand them a task it never mentions, say "draft a tweet about our new feature", and you can already predict the result: plain voice, no buzzwords (so no "unlock"), a source if it states a number, and a question back if the brief is thin. That predictability is the whole value. The most common weak spot is the "never" section being empty or polite. Vague nevers ("do not be unprofessional") do nothing. Sharp nevers ("never invent a statistic", "never use the word leverage") are enforceable and they are where your taste actually lives. If you let Claude draft it, do not accept the draft whole. The draft is a starting point. You are the one who knows which rules are real, so you make the final calls.
WATCH FOR
✗You start from a blank page. Start from your personal constitution. The project inherits you, then adds its own rules.
✗Your "never" list is vague or empty. Vague nevers do nothing. Sharp, specific ones are where your taste gets enforced.
✗You accept Claude’s draft without editing. The draft is raw material. You decide which rules are actually yours.
✗You write it once and never revisit it. When the project misbehaves the same way twice, add a rule. The file grows with use.
WHAT YOU LEARNED
The takeaways
CLAUDE.md is the file a project reads first and treats as its standing law. It is the project’s constitution.
Writing one moves you from prompting better to building the world the AI works inside: rules set once, applied to every answer.
The fastest start is to inherit from your personal constitution, then add what this project specifically needs.
You can have Claude draft it with you, but you make the final calls on which rules are real.
Your project · step seven
Your thread project now has a CLAUDE.md: standing rules it follows on every task, seeded from your constitution. This is the moment your project stops being a chat and becomes a thing you built. From here, when it gets something wrong twice, you do not re-explain in the chat. You add a line to the file. Next chapter: teaching it a repeatable job with a Skill.
A prompt is something you say. A CLAUDE.md is something you build. The day you write your first one is the day you stop being a tourist.