The background you should stop retyping

The same eight facts, every time

If you use a model for work, roughly the same background is missing from every prompt you write. Your role. Your country. Your currency. The tools your team uses. The three words your organisation uses in a non-standard way. The audience you write for. The things you are not allowed to say.

Most people supply these ad hoc, which means they supply them inconsistently and often not at all.

Write them once, as a block you paste or store. Ten minutes, and it improves every prompt you write afterwards.

What actually belongs in it

Keep it to things that change the right answer. A worked example, for a bookkeeper in Nairobi:

Who I am: bookkeeper at a 14-person building-materials wholesaler
in Nairobi. Not an accountant; I work with one.
Where: Kenya. Currency KES. Financial year ends 31 December.
Tools: QuickBooks Online, Google Sheets, M-Pesa for most receipts.
Our words: "job" means a customer order, not a task.
  "Credit" in our sheets means money owed TO us, not by us.
Audience: I write for the two directors, who read on a phone.
Constraints: I never share customer names or phone numbers.
Style: short. No preamble. British spelling.
Today's date: I will state it when it matters.

Nine lines. Every one of them changes an answer: the currency line alone prevents a monthly irritation, and the house-vocabulary lines prevent a class of error that is hard to spot, because a confident answer using the standard meaning of “credit” reads perfectly well. Note what is not in there: nothing about being a hard worker who values clarity, no instruction to be an expert assistant, no motivational preamble. Those change nothing.

Where to put it

Custom instructions or a project setting, if your tool has them: ChatGPT’s custom instructions, Claude’s Projects, Gemini’s Gems, a Copilot agent. These prepend your block to every conversation automatically. This is the best option when the tool offers it, with one caveat below.

A note you paste, if it does not. Keep it in a plain text file or a phone note. Pasting nine lines takes four seconds.

At the top of the prompt, when the task is one-off and important enough to be worth the four seconds even though the tool has no settings page. A brief pasted by hand is a brief you can see, which is occasionally an advantage over one applied invisibly.

A separate brief per area of work, once one block starts contradicting itself. A bookkeeper who also runs the company’s social media should have two. Trying to serve both from one file produces a brief full of “when I am doing X…”, which is a brief that has outgrown itself.

The caveat that catches people

A standing brief applies to everything, including the times it is wrong.

Somebody sets “always answer in bullet points, never more than 200 words” and then wonders why every explanation is thin and every draft arrives as fragments. The instruction was fine for status updates and wrong for the other eighty per cent.

Two rules keep this from happening:

Put facts in the standing brief; put style in the prompt. Your currency does not change between tasks. Your preferred output shape changes constantly.

Re-read it monthly, and after anything surprising. If an answer arrives oddly clipped, oddly formal, or oddly obsessed with a topic you mentioned once, the standing brief is the first suspect: precisely because you cannot see it in the conversation.

Distinguish it from memory

Several products now save facts about you automatically as you chat. That is not the same thing, and it deserves its own lesson, which is next. The difference in one line: a standing brief is a document you wrote and can read; memory is a set of notes something else took about you.

What it is worth

The gain is not that any single answer becomes brilliant. It is that the floor rises. You stop getting American spelling, dollar amounts, advice that assumes an accountant, and explanations pitched at somebody who does not know what your company does.

That is most of the gap between people who say these tools are useful and people who say they are generic: not a better prompt, but the same prompt sent by somebody the model actually knows something about.