Write your AI custom instructions by letting it interview you, once

Watch what an assistant does with a title. Tell it you are a finance director who likes things concise, and it answers the average finance director: it explains what a forecast is to someone who signs forecasts off, and pitches its recommendations below the level you work at. The model fills the gap you left with the most likely version of your role, which is rarely you.

A title leaves out the three things that would change the answer: what you decide, what you already know and how you want the answer laid out. Two finance directors with the same title can differ on all three. One owns the budget and wants the recommendation in a line; the other advises a board and wants the working shown. The assistant cannot tell them apart unless somebody writes it down.

Writing it down yourself is harder than it sounds. Describing your own job to an empty settings box is awkward, and you have no way of knowing which details a model needs. Answering questions is easy, so turn it round: let the assistant ask, once, and keep what it writes.

Start a new chat in your work account and paste this as it stands. It is the prompt from the film, and every clause in it does a job, which the next section explains.

Answer in full sentences and give a recent example where you have one. A decision you made last week tells it more than a description of your remit. If a question goes somewhere you would rather it did not, say skip and ask for the next one.

When the questions stop, it writes the page. In the film the page is headed “How to work for me”, which is a good title to ask for if yours comes back without one.

The interview moves the effort. You no longer need to know what a model has to hear about you, because it asks, and each answer shapes the next question. Both vendors recommend the same move for other work: OpenAI’s Codex guide says that when you only have a rough idea, you should ask the assistant to interview you and challenge your assumptions before it writes anything [9], and Anthropic documents an interview pattern for larger pieces of Claude Code work [10].

Here is what each part of the prompt does.

Anthropic’s course on Cowork lands in nearly the same place. It says a first standing brief for Claude needs three things: who you are and what you do, the shorthand you use, and how you like output delivered, with the corrections you keep repeating as the next candidates [5]. The prompt covers the first and the last directly. If your team runs on acronyms and code names, add a line asking it to note your shorthand too.

ChatGPT’s custom instructions hold 1,500 characters on the Free and Go plans and 5,000 on Plus, Pro, Business, Enterprise and Education [6]. A full page will not fit in 1,500, so on those two plans ask for less than the limit and check the length before you paste, because models count characters loosely.

If something has to go, keep what you already know and how you like answers, and cut what you decide down to a line. Those two change every reply; your authority mostly matters when you ask for a recommendation.

Roles drift. A new remit, a new team or a new boss makes half the page wrong, and a wrong page is worse than none, because the assistant keeps applying it. Run the interview again against the page you already have, so it asks only about what changed.

Read the list of changes before you paste the new page over the old one, and keep the old copy with its date, so you can see what moved.

Each product has one setting that applies across your chats, and that is where the page belongs. Each also has settings that sit above or beside it, and those decide whether your page is used at all.

The Claude column is from Anthropic’s pages on personalisation [1], on Cowork [2] and on memory [4], and the ChatGPT column from OpenAI’s pages on custom instructions [6], projects [7] and memory [8].

Claude has one more layer, above your own. On Team and Enterprise, an Owner can set Organization instructions that Claude follows in every conversation across the organisation, up to 3,000 characters, and where a personal instruction contradicts one, Claude favours the organisation’s [3]. Only Owners and above can read them [3]. If your company uses them, your page does not need to repeat what they say, and it should not argue with them.

The page guides the assistant; it does not bind it. OpenAI’s own wording is that ChatGPT considers custom instructions across your chats [6], and Anthropic says even its organisation instructions work at the level of the prompt rather than as hard enforcement [3]. Expect the page to shift the defaults, and check the first few answers after you save it. Anything that has to happen every time belongs in a control the system enforces rather than a sentence in a settings box.

Treat what comes back as a draft. The assistant wrote it from your answers, so it will sometimes turn one example into a rule, or record as a preference something you mentioned once. Read it as you would a brief from a new colleague, and correct what it guessed.

Interviews pick up more than you meant to give. Answers about work drift into health, family and opinions of the people around you, and the write-up can carry all of it. Claude’s memory leaves sensitive topics such as health out by default unless you opt in [4], but the instructions page is text you paste yourself, so the filter is you. Take out anything you would not show a colleague before you save it.

Use the work account. The answers describe your employer, its gaps and its people, which makes them company information, so run the interview only in the account your company has approved and keep confidential detail out of any personal or consumer account. If your company has not chosen an assistant yet, our guide to how the Claude and ChatGPT business plans compare sets out what each plan publishes on admin controls and data location.

Other settings can outrank the page without telling you. A ChatGPT project’s instructions override it inside that project, and in Enterprise and Edu workspaces personal custom instructions are not available in projects at all [7]. If answers inside one project ignore your page, check the project first.

Keep procedures off the page. Anthropic’s guidance is that general rules go in instructions and a repeatable process goes in a skill [14], and a month-end routine pasted into your instructions crowds out the lines about you. And the page goes stale as your job changes, which is what the update prompt above is for.

The page the interview writes is one person’s half of the job. The company’s half decides what everyone’s assistant starts from: the account people sign in to, whether members can turn memory on, what any organisation instructions say and which shared skills are there on the first day. Setting up and rolling out these tools for companies is core 1AYM work, including the instructions, skills and permissions around them, and our engagement file on AI platform enablement at a large international marketing agency shows what re-architecting a skills estate on a production platform changed.

We would build that company half under our AI governance implementation work, so the rules sit in settings an owner controls rather than in a memo each person has to remember. It can start as a small fixed-scope statement of work. If a second opinion on your setup would help, book a call or email enquiries@1aym.com.