Understanding AI

The brief is your agent's personality

One document, read before every message you send. Most people never write it.

Write Yours Now

This is the whole job. Paste this into Claude, ChatGPT, or Claude Code exactly as written and answer its questions. It asks which app you use first, interviews you one question at a time, then hands back a finished brief ready to paste into your custom instructions, your personality box, or a CLAUDE.md — and tells you exactly where that is.

Paste this into your AI
I want you to write my AI brief with me. The brief is the standing document you read before every message I send, the thing that sets your personality and your rules. I want to walk away with a finished version I can paste straight into the right box. First question, before anything else, ask me which AI I use day to day. Claude, ChatGPT, or Claude Code. That decides where this document ends up living. Then interview me one question at a time. Don't write anything until you've finished asking. Work through all five of these: 1. Who you are. The role I'm hiring you into, and what you're there to do for me. 2. How you behave. Tone, how you speak to me, how formal, how long your answers run, British or American English. 3. My standing rules. The way things are always done here, whatever the task is. 4. The never dos. The hard lines you are not allowed to cross, ever. 5. When you come to me. The decisions I don't want you making on your own. Ask about the things I probably haven't thought of. If I give you a vague answer, push back and ask for the specific version. If a rule only applies sometimes, it doesn't go in. Only the always rules belong in the brief. When you have all five, write the brief as one finished, paste-ready document. Address it to the AI, "you are", "you always", "you never". Use my words where I gave them, keep it plain English, short enough to read in under a minute, and leave nothing for me to fill in. Then tell me exactly where to paste it for the app I named at the start. For Claude, a Project's instructions, or the personal preferences box in Settings if it should apply to everything. For ChatGPT, a Project's instructions, or Custom Instructions in Settings. For Claude Code, save it as a CLAUDE.md file in my project folder, and create that file for me rather than telling me how. Finish by reminding me this is a living document. Every time you guess something wrong, I come back and we add the rule that stops it happening again. That is how the personality gets built. One question at a time.

Where To Put What It Writes

The interview ends by telling you this for your app, but here it is up front. The only decision is whether the brief is for one kind of work or for everything you do.

In Claude. For one kind of work, make a Project and paste it into that Project's instructions — every chat inside the Project reads it first, chats outside are unaffected. For everything you do, it goes in the personal preferences box in Settings.

In ChatGPT. Same shape. A Project's instructions for one kind of work, a custom GPT's instructions if you want to hand it to colleagues, or Custom Instructions in Settings for everything.

In Claude Code. The brief is a file called CLAUDE.md sitting in your project folder, read at the start of every session — the interview prompt creates it for you. If you build, the deeper guide is the three tiers of CLAUDE.md.

Anything else your work gave you, Copilot or an in house tool, the names differ and the pattern doesn't. Somewhere in it there is a box labelled instructions, and this is what that box is for. These menus get rearranged every few months — you are looking for the box that describes how the AI should behave, not what you want it to do today.

Why It Works

Every time you send an AI agent a message, it reads one document before it reads your message. Every time. Not the first time and then from memory, every single time you hit send.

That document is its brief, and it is basically its personality. It is the reason one AI comes back curt and useful and another comes back with three paragraphs of throat clearing, when they are the same AI underneath.

Here is the bit people get wrong. The brief is not the task. The task changes every day, and you type that into the chat like you always have. The brief is who this person is and how they behave, written down once. A handbook, not a job ticket. You would never hand a new starter the handbook fresh each morning, and you would never expect the job ticket to explain how the company works.

Because it gets read before every message, whatever is in it holds. Written down means your AI shows up the same way every day. Same tone, same rules, same lines it will not cross, whether it is Monday morning or the fortieth message of the afternoon.

Unwritten means it improvises. And this is the failure mode nobody warns you about, because it does not improvise nervously. It fills every unwritten rule with a confident guess. It picks a tone, decides how formal to be, decides whether that is a decision it can make on its own, and it tells you none of it. Like a new starter with no handbook who is far too polite to ask, so they just get on with it and you find out three weeks later.

The Five Things That Go In It

This is what the interview above walks you through, five things, a few lines each, in the words you would use if you were sitting next to someone explaining the job. If you would rather write yours by hand, every example below is built to be adapted, so change the nouns and it is yours.

1

Who they are

The role you are hiring them into, and what they are there to do for you. Nobody starts a job knowing what the job is, and an AI with no role picks one for you, usually the most generic version it can find. Say it plainly. Something like "you are my research assistant. I bring you a company name and you come back with what they do, who runs it, and anything recent I should know before I get on a call with them."

2

How they behave

Tone, how they talk to you, how formal, how long the answers run. This is the difference between the colleague who sends you three lines and the one who sends you a five paragraph essay with a preamble. Both know the answer. Only one of them is pleasant to work with. Something like "talk to me like a colleague, not a consultant. Short answers. No preamble, no summarising my question back to me, get straight to the point."

3

Your standing rules

The way things are always done here, no matter what the task is. Every workplace has these and nobody writes them down, which is exactly why a new starter breaks them in week one. These are the ones that apply on every single job, not the ones that apply to Tuesday's job. Something like "British spelling, always. Money in pounds. When you give me a fact, tell me where it came from. If a document has a date on it, tell me the date."

4

The never dos

The hard lines. Not preferences, not nudges, the things you would have a serious conversation about if a person did them. Being blunt here costs you nothing and saves you the one bad day. Something like "never promise a delivery date. Never quote a price that isn't in the price list. Never put a customer's name in anything that leaves this chat."

5

When they come to you

The decisions you don't want it making on its own. A good new starter knows which calls are theirs and which ones get walked down the corridor first, and they only know it because someone told them. Without this line the AI decides for itself, and it will nearly always decide it can handle it. Something like "if it involves a refund, a complaint, or anything a lawyer would care about, stop and ask me before you do anything else."

One Honest Bit

Your first brief will not be right. Not because you wrote it badly, but because you cannot possibly list every rule in your head before anything has gone wrong. Half of them you do not know you have until something breaks one.

So treat it as a living document. Every time it guesses wrong, that is the AI telling you exactly which rule was missing. You go back, you add that one line, and it never guesses wrong on that thing again. Do it for a fortnight and you have a brief that is genuinely specific to you, built out of real mistakes rather than imagined ones. That is how the personality actually gets built.

And the honest version of the promise. A brief reduces the guessing, it does not eliminate it. There will always be a situation you did not think to write a rule for, and it will still make a confident call in the moment. The point is that the list of things it has to guess at gets shorter every week, instead of staying exactly as long as it was on day one.

Where This Goes Next

If you have landed here without the overview, an AI agent is a chat with a job is what an agent actually is, and the brief is one of the three things it is made of. Teach AI to do a task your way is the next step up, saving one specific job so you never re-explain it. And custom instructions is the everyday version of the same idea, standing instructions for your own normal chats rather than for an agent.

Part two of the series is tools are how your AI stops being useless, the connections that let an agent actually do the job rather than describe it. Which apps to connect first, the ones I would never connect, and why read only is a real option.

Part three is the context is what your agent actually knows, the background file that stops every answer being written for anyone. A brief tells it how to behave, the context tells it who it is behaving for.