Your profile in Claude Code

Third conversation this week, and every time the same opening: that you run an accounting practice, that you work alone, that you write to clients formally rather than casually. Claude Code can save some of that itself, in auto memory – but with no guarantee: the notes are unstructured, and whether this particular one gets used is something AI judges on the fly, from a single line in MEMORY.md’s index. Sometimes it lands, sometimes it doesn’t, and you explain everything again from scratch.

This lesson turns that into one file: profile.md. You’ll answer a series of questions about yourself and your work, Claude Code will build your profile from them, and then reach for it on its own whenever the conversation is about you or your work.

For the exercise you need a folder you already work in, with a CLAUDE.md file inside. If you don’t have one, start with the lesson on your first CLAUDE.md file and come back here.

A profile lives in its own file, not in CLAUDE.md

The first temptation is to add the profile to the bottom of CLAUDE.md. Better not to give in. That file, alongside other instructions, should hold only the basics about you, something like a business card: your name, how you’d like to be addressed, contact details if the assistant is meant to use them day to day. Your profile answers a different question entirely – who you are, how you work and where you’re heading – each of those threads spelled out in far more depth.

Claude Code’s documentation recommends keeping CLAUDE.md under two hundred lines, so packing all of that in there could push the file past what’s recommended. There’s a second reason too, one that matters more day to day: CLAUDE.md loads in full at the start of every conversation, and your profile is only needed in some of them.

The split we use ourselves and in our workshops: CLAUDE.md is your business card and your work rules. profile.md is your full documentation.

Don’t write the profile yourself – let Claude interview you

Writing a profile from scratch is hard for the same reason a short bio about yourself is hard: you don’t know where to start or what matters. So today you don’t write it at all. You ask Claude Code to interview you, and you answer the questions one by one.

Run Claude Code in your folder and paste this in:

Interview me and write a profile.md file from it – my profile for you.
Ask one question at a time and wait for my answer before asking the next.
Start with who I am and what I do, then ask what takes up most of my time, and finish with what my plans are.
Write the file only at the end.

In our workshops this interview runs to eighteen questions and takes about half an hour, working through the answers included. That’s not time wasted: it’s the one moment you lay everything out at once.

Answer briefly and plainly, in full sentences. If a question comes up you don’t have an answer to, say so outright – the same way you would with a person.

Three parts of the interview: who you are, what blocks you, where you’re heading

Who you are. What you do day to day. What tools you use – email, calendar, invoicing software, a CRM. What your typical working day looks like and when you have your best hours for focus.

What blocks you. Which tasks eat the most time even though they shouldn’t. What frustrates you most about this work. What’s on your plate right now – projects, clients, commitments. What you decide alone, and what needs settling with someone else. This part is uncomfortable, and that’s why it works: you’re describing your work as it actually is.

Where you’re heading. Goals for the coming year. What you want to do less of, and more of. Whether you’d rather get feedback right away or gathered at the end.

None of this has to land in the file. Revenue figures, client data, anything you wouldn’t tell a new hire on their first day – leave it out of the profile for now, you can come back to it once it makes sense to.

Boundaries: what Claude Code should ask before it acts

The interview ends with the most practical question of all: what the AI assistant should know so it doesn’t get in your way. This is where you write down boundaries – situations where it should stop and ask before going further. If that question doesn’t come up in your interview, just dictate the boundaries at the end, before Claude Code writes the file.

Try to phrase these as positive guidance rather than prohibitions. The reason is simple: “never look in the archive folder” gives away the existence of a folder Claude shouldn’t have to think about at all.

Three examples to adapt into your own, because a boundary has to match what genuinely needs your say-so:

• “Show me the email to a client for approval before you send it.”
• “Only change anything in the contracts/ folder once I’ve agreed to it.”
• “Let me read any text meant for publication before it goes out.”

Claude Code follows these sentences, but nothing forces it to – it’s an instruction, not a technical block. What decides whether it stops to ask before changing a file on its own is the permission mode, and which mode lets through what is covered in the tip on permission modes. Boundaries from a profile reach further than modes do: they cover sending an email, publishing a text, reaching out to a client – moments where nothing on your disk changes at all, and you’d still rather see what came out first.

Point to the profile in CLAUDE.md, instead of pasting it in

The profile.md file sits in the folder, but Claude Code won’t reach for it on its own – it doesn’t know yet that it exists, or why. One sentence in CLAUDE.md is enough to point at it:

Add a sentence to CLAUDE.md: my profile is in profile.md, read it whenever you need to know who I am and how I work.

That’s the whole thing – no pasting the profile in, no import. Thanks to that, the profile pointed at in the sentence sits waiting on disk and AI only opens it when it actually needs it – usually within the first minute of a conversation about your work, not when you ask about today’s weather.

Several profiles, each in its own folder

Does one profile have to carry your whole life? The same person runs a business, and in the evening writes fantasy stories or helps the kids with maths homework (a subject they were never especially good at themselves) – and those are two different contexts, where different things matter, different things are allowed, and the writing sounds different.

So build a second profile with the same interview as the first, just run in a different folder, and save it there. Claude Code reads CLAUDE.md from the folder you run it in, and from the directories above it. So placing profiles correctly in your folder structure means only one of them ever gets read. Run Claude Code in your work folder and it reaches for the work profile; run it in a personal folder and it reaches for the other one. You don’t switch anything by hand – you just need to be in the right folder.

Each of those folders needs its own CLAUDE.md with its own sentence about the profile. Two profiles in one folder don’t give you a choice: Claude Code will see both and try to reconcile one with the other. And when you run Claude Code somewhere else entirely, it won’t see either – which is fine, because wherever that is, you’re doing something that has nothing to do with your work.

Check in a new chat that the profile actually works

Close Claude Code and run claude again. In the previous conversation the profile was known because it had just been written before your eyes – that proves nothing. What counts is a new chat, and a question that has no good answer without the profile:

Who am I? What do I do? What's my working style?

A precise answer means everything’s in place. A vague one, or a question back like “so what do you do?”, means the profile wasn’t loaded: check whether the sentence pointing to it is really in CLAUDE.md, and whether the file’s name matches that sentence exactly. You can also ask outright why Claude Code didn’t take it into account – a question worth keeping for any instruction that doesn’t work.

Update the profile when your work changes

The profile describes you as of the day it was written. A new service comes along, your client base shifts, a big project that ate half your week wraps up, your working hours change. You don’t open the file for this either – Claude Code writes every change, down to a single line, the same way it did for the interview. Tell it about the change in conversation:

Update profile.md: I now also work with small online shops, not just accounting practices.

A simple habit for this: whenever a change is big enough that you’ll be explaining it to a client or someone in your family anyway, tell Claude Code the same thing.

What to take away from this lesson

Your profile gets its own file: profile.md. CLAUDE.md stays your business card and your work rules, profile.md is your full documentation – and the second one is useful outside the terminal too.

A profile comes out of talking about yourself. Claude Code runs the interview, you answer the questions one by one, and it writes the file at the end from your answers.

Boundaries are part of the profile, phrased as a request to ask. “Show me the email before you send it” works better than “don’t send anything on your own”, and they reach further than a permission mode, because they also cover what changes nothing on disk at all.

CLAUDE.md holds a sentence about the profile, not the profile itself. That way Claude Code opens it when it’s actually needed, instead of pulling it into every conversation – and separate folders with separate profiles give you a work version and a personal one without switching anything.

Tasks

Tick these off as you go. The state is remembered in your browser, so you can come back to this list tomorrow.

• Interview done, profile.md written from the answers
• Profile covers who you are, what you do, and how you work
• Boundaries added: situations where Claude Code should ask first
• Sentence about the profile added to CLAUDE.md
• New chat opened, check question asked
• I know what to do if the answer in the new chat came out vague