Hey friends,
You know that feeling when you open a new AI chat, start explaining your project, and somewhere around the third paragraph you think… didn’t I already do this yesterday?
AI is smart.
That’s never really been the issue.
The annoying part is how fast it forgets the stuff that actually makes the work good.
The things that actually matter like your rules, your audience, your tone, what you’re even trying to make.
So every new session starts from zero.
And before you can get to the actual work, you’re back there rebuilding the same context all over again.
And if you’ve started using AI agents, like Claude Code, Claude Cowork or Codex, this matters even more.
This is also the kind of practical setup I’m teaching inside my upcoming course, Learn Claude Cowork in Cowork. If you want first access when it opens, you can join the waitlist here.
These tools go further than answering questions. They take actions on your behalf.
They write, build, browse, and run tasks for you. The more context they have before they start, the better that work gets. And the less you end up correcting.
That’s why this episode matters.
Because the difference between generic AI and actually useful AI is often not the model.
It’s the context.
And there’s a file that handles that part for you.
It’s called CLAUDE.md.
Let me show you exactly what it is.
📌 What Is CLAUDE.md?
Earlier this week I virtually attended the Claude Code Aus Community and I took this screenshot to share with you
CLAUDE.md is a plain text file that Claude reads at the very start of every session. Think of it as the note you leave before the work begins, so Claude arrives knowing the room.
You write it in normal language. No code or technical setup.
Just plain instructions about your project, your preferences, and what you need.
Here’s what makes it useful:
- Persistent: Claude reads it fresh every session, automatically
- Plain language: written the same way you’d brief a new assistant
- Flexible: works for a specific project, or across everything you do
- Easy to start: one command generates a starter version for you
One thing that’s easy to miss is this doesn’t have to be one giant file for everything.
You can keep project context at the project level, and keep your personal preferences in a separate file that follows you across projects.
That way Claude learns what’s specific to the work in front of it, without mixing it up with how you like to work everywhere else.
Using something other than Claude?
Outside of Claude, you’ll often see the same idea show up as a file called AGENTS.md.
✨ How I Used It
✅ 1. Set My Project Up Once
Last year, every time I started a new session to work on my newsletter, I’d spend the first few minutes catching Claude or ChatGPT up. Who my audience is, what the tone is, what I’m building, what to avoid.
It felt productive but it wasn’t.
I was just rebuilding the same starting line every single time.
Once I wrote all of that into CLAUDE.md, that setup disappeared. Claude just… knew. The session started at the actual work.
What I learned pretty quickly is this file works best for the things that stay true across sessions.
Your audience, tone, rules and what good looks like. Not the random one off task you need done today. That still belongs in the chat.
Pro tip: Write your CLAUDE.md like you’re briefing someone on their first day. Assume they’re sharp but know nothing about your project yet.
✅ 2. Lock In How I Like Things Done
Generic AI agents are easy to spot.
They technically works.
But it makes choices you would never make.
Uses the wrong tool. Takes too many steps. Asks at the wrong time.
I added a short working rules section to my CLAUDE.md describing how I like tasks handled, when to ask before acting, what tools to prefer, and what to avoid.
The difference in output was instant.
Claude stopped freestyling and started working in a way that actually fit how I work 😅
Pro tip: Add a few simple rules like “ask before sending anything”, or “show the plan before making changes.”
✅ 3. Define What “Done” Looks Like
The other thing that drains time is correcting the same mistakes across sessions.
The app works, but the flow is messy. The automation runs, but there’s no fallback. The agent finishes the task, but leaves out the one thing you actually needed.
I added a short “done properly” section listing what a finished result needs before Claude hands the work back.
- What should exist at the end
- What needs to be checked
- What to include with the final output
Claude stopped guessing and started landing much closer to the mark on the first pass.
Pro tip: Don’t just describe the task. Describe the finish line. Tell it what has to be true before the work counts as done.
🛠 How to Get Started (3 Steps)
- Open Claude Code and type /init: Claude will look at your project and generate a starter CLAUDE.md for you (in the desktop app just start a new session and prompt Claude to update the file)
- Edit it in plain language: Fill in what matters. Your project, your audience, your tone, what to avoid.
- Add your “done properly” definition: Tell Claude what a finished output looks like for you. Format, length, structure, and any non-negotiables
A note on length: Keep the file short. Research across 2,500+ real projects found that under 150 lines is the sweet spot. Every line that doesn’t need to be there actually dilutes the ones that do. For each line you write, ask yourself: would Claude make a mistake without this? If the answer is no, cut it.
Pro tip: Let it improve over time. When Claude gets something wrong, add that correction to the file and tell Claude to update it periodically with learnings about you.
📈 Why This Matters
It’s easy to get a result from AI. Open a chat, type a prompt, something comes back.
Then people wonder why the quality swings so much from one session to the next.
A better prompt can help. But this goes deeper than prompting.
It’s a context problem.
The quality of the work depends on what the tool understands before the work begins.
What changes things is giving AI a place to hold the parts that should not need repeating. How you work. What you’re building. What matters. What “done properly” means.
CLAUDE.md (and AGENTS.md, for everything else) is that place.
It takes the setup work out of your head and puts it somewhere the tool can actually use every time.
If this kind of workflow is what you want to get better at, that’s exactly what I’m building inside Learn Claude Cowork in Cowork. You can join the waitlist here if you want early access and the early-bird price.
🫡 Final Thoughts
I know CLAUDE.md sounds technical. That’s exactly why I wanted to write this.
Because the problem it solves is not technical at all:
The drag of having to re-explain yourself every time you sit down to work with AI.
Whether you’re using Claude Code, Cowork, Codex, or something else, start with a context file. Give the tool a clear starting point before it begins.
Try it this week and you’ll probably feel the difference straight away.
If this one landed, forward it to someone who’s been saying AI still doesn’t really get how they work. This is one of the simplest ways to fix that.
To AI that actually knows you,
Nahid
