Custom Instructions, GPTs, and Projects: The Practical Guide

If you still type "I'm an accountant, reply in English, keep it short, skip the pleasantries" at the start of every conversation, you're doing by hand what your tool can do on its own. These sentences belong in a permanent setting: the AI reads them before your first message, every time.

The vocabulary varies from tool to tool and menus change often, but the idea stays the same everywhere. There are instructions that are always read, separate from what you type, and they steer all the answers.

Four levels, from preference to dedicated assistant

Not everyone needs to build a custom assistant. The following levels differ in scope. Names and availability by plan change: check in the tool as you read this.

When in doubt, start with the first level. You'll move to a project the day you have reference documents to attach to it, and to a custom assistant when a task comes up often enough to deserve its own tool.

What goes in the instructions, what stays in the prompt

The rule comes down to one question: will this still be true in three months and for almost all of my requests? If yes, it goes in the instructions. If not, it stays in the day's prompt.

What works well in permanent instructions: your profession and level, your formatting habits, the mistakes you never want to see again, what the AI should do when it's unsure. What weighs them down: the details of a current file, a long list of rules you only apply once a month, generalities like "be helpful and accurate."

Compare these two versions for a freelance accountant.

You are an expert assistant, highly competent and professional. Always give the best possible answers. Be precise and helpful.

This text gives no usable information. A model is already trying to be helpful, and "expert" doesn't tell it in what field or for whom.

I'm a self-employed accountant in France, working mostly with tradespeople and very small businesses. [ADAPT TO YOUR PROFESSION AND CLIENTELE]

For my requests:
- Reply in English, with no introduction or pleasantries. Start with the answer.
- When you write for a client in my place, use simple sentences and avoid accounting jargon that isn't explained.
- For a tax or social security rule, tell me what may have changed and where to check an official source. Never present a rate or a threshold as certain.
- If a piece of information is missing to answer correctly, ask a question before answering, two questions maximum.
- If I ask for your opinion, give it and tell me what would change your mind, without an exhaustive list of pros and cons.

Each line changes an observable behavior. That's the test to apply to all your instructions: if you remove the line and the answer doesn't change, delete it.

Build an assistant dedicated to one task

A custom assistant, or a project, becomes interesting when it combines a single mission with documents. Take a craft business, a woodworking shop for instance, that answers quote requests by email every week. It can upload its price list of services and two or three template replies, then give it these instructions.

Mission: you help [COMPANY NAME], [BUSINESS] in [CITY], reply to quote requests received by email. You handle only this task.

How it works:
1. I paste in a customer request.
2. You list in a few lines what the customer wants, and the information missing to price it.
3. You write a reply ready for review, under 150 words, addressing the customer formally, in the tone of the examples provided in the documents.

Rules:
- Rely only on the "Services" document for timelines and the services offered. If the request falls outside that scope, tell me.
- Never state a firm price. Write [PRICE TO BE CONFIRMED] instead.
- Don't state an installation date unless I've given you one.
- If the request isn't a quote request, simply say so and don't write anything.

The two prohibitions on price and date are the most important in the text: they're the two mistakes that would cost dearly. Always write down the limits that protect your business, and put them in short sentences.

If you already know the elements of a good prompt, the structure will look familiar. The difference is that this text has to hold up for requests you haven't seen yet.

The pitfalls of permanent instructions

They're not a guarantee. A model can ignore a rule, especially if the list is long or contradictory. For a limit that really matters, always reread the output.

They can be forgotten. If an account setting imposes a style, you may be baffled by strange results three weeks later. When an answer surprises you, remember to check your instructions and the tool's memory if it's turned on.

They're not a vault. Don't put in passwords, API keys, or sensitive client data. A shared assistant can be pushed into revealing its instructions or the contents of its documents. To know what's better kept out of these tools, read what not to paste into a chatbot, and check the terms of your plan, especially if you handle client data.

Write your instructions from what you correct

Open your last five conversations and note what you had to ask for again: "shorter," "no headings," "in English," "be more formal." Those corrections are your first set of instructions. Paste them into your account setting, then ask the AI a question you ask it often and see whether you still have something to fix. If you want to test these instructions more seriously, keep three typical requests and replay them after each change.