Your First Custom GPT: A One-Afternoon Build With No Code

People keep telling you to build a custom GPT as though that were a sentence with an obvious meaning.

If you have read it three times this month and still have no picture in your head of what you would actually be doing, you are not the only one. The name is doing you no favours. It sounds like programming and it is not.

Here is what it really is, and then a block you can fill in and paste, which is the part that decides whether the thing is any good.

What you are actually making.

It is the same ChatGPT you already use, with three things attached to it permanently.

A set of standing instructions it reads before every single answer, so you never have to explain who it is or what you want again.

Files it can look things up in. Your policy, your price list, the twelve questions you get asked constantly.

A choice of what it is allowed to do, such as searching the web or making images, switched on or off.

That is all. No code, no account beyond the one you have, and it lives in your ChatGPT sidebar afterwards like an app you made.

I want to be straight about one thing early, because the word agent gets thrown at this constantly. A custom GPT is not an agent. It is level two, tidied up: it still waits for you, and it still cannot do anything outside the chat window. That is not a criticism. It is why you can finish one this afternoon.

The four steps, and only one of them is hard.

01 · Say what it is for. In plain English. An assistant that answers questions about our return policy.

02 · Give it the knowledge. Upload the documents. A one-page policy is enough to start. You do not need a library.

03 · Choose what it may do. For most first builds, turn everything off. You want it answering from your files, not wandering off to search the web.

04 · Test it and correct it. Ask it the awkward questions. Tell it what to do differently, in ordinary language, and watch it change.

Three of those take about ten minutes between them. The one that decides everything is the instructions, and that is where nearly every disappointing custom GPT goes wrong.

The free AI Agents pack. The Agent-Plus Prompt Builder in there writes instruction blocks like the one below for you, if you would rather start from something than from an empty box. It comes with over two thousand ready-made n8n workflows, a workbook on the twenty building blocks that cover most of them, and the checklists I run through before letting anything work on its own. Download the AI Agents pack →

You already know six letters. This needs one more.

If you have used my six-part framework, you know the shape: context, role, objective, format, tone, conditions. Six things that turn a hopeful request into a repeatable one.

For something that runs on its own, there is a seventh, and it goes at the front.

A is for agent, and it means how should you behave. Here is why it matters. When you are sitting in a conversation, you can see a bad answer and correct it. When somebody else is using the thing you built, at nine on a Tuesday, you are not there. So you have to decide in advance what it does when the question is unclear, when the information is missing, when somebody asks it something it was never meant to handle.

Written out in full it is Agent, Context, Role, Objective, Format, Tone, Conditions. Seven lines. That is the entire framework for this.

The block to fill in and paste.

Copy this into the Instructions box and replace the parts in brackets. Keep the headings. They are what stops it drifting.

AGENT Answer immediately. Do not ask follow-up questions. If something needed is missing, say what is missing rather than guessing. CONTEXT [One or two sentences about the business. We are a small shop selling handmade leather bags, online and in one store.] ROLE You are [what it is. a calm, precise customer service assistant]. OBJECTIVE Your only job is to [the one thing. answer customer questions about returns, delivery and product care, using the uploaded documents]. FORMAT Answer in [shape. two or three short paragraphs, no bullet points, no headings]. TONE [How it should sound. warm and plain, the way one person explains something to another]. CONDITIONS Always [what must be there. quote the relevant rule from the documents]. Never [what must not. invent a timeframe, a price, or an exception]. If the answer is not in the uploaded files, say so plainly and offer to pass the question to a person.

That last line under Conditions is the most valuable sentence in the whole block. Without it, a question your files do not cover produces a confident, fluent, completely invented answer.

The same assistant, before and after.

Before. Instructions read: You are a helpful assistant for our shop. Answer customer questions. Somebody asks whether they can return a sale item. It gives them a friendly, general, plausible answer about most retailers usually allowing fourteen days. Your actual policy says something else.

After. With the block above, the same question returns the rule from your document, in your voice, in the shape you asked for. And when somebody asks about international delivery, which is not in your files, it says so instead of improvising.

Same model, same afternoon, same files. The difference is seven headings.

What it will not do.

It cannot act. It writes an email, it does not send one. It can tell somebody the return process, it cannot start one. Making it do things is possible and it is a different afternoon.

It does not learn from conversations. If you want it to know something new, put it in the instructions or upload a file. Correcting it in a chat improves that chat and nothing else.

It is not a private vault. Content from the files you upload can appear in answers, so take out anything you would not want quoted back at you: personal details, internal notes, anything you are contractually holding. Clean the file first, upload it second.

There is a ceiling on files. Twenty per assistant. That is more than a first build needs and less than a whole company archive, which is a different job with a different answer.

One small thing to do before the next post.

Do not open anything yet.

Write one sentence, on paper, that starts Your only job is to.

If you cannot finish that sentence in one line, the assistant will not work, and no amount of building will rescue it. If you can, the rest genuinely is an afternoon, and most of the afternoon is you asking it awkward questions and watching it get better.

If you want the whole thing in one place.

If you would rather have the complete system than one piece at a time, that is what I wrote AI Agents Bible for. Five books in one, with Custom GPTs, Make.com and n8n each built step by step, thirty worked examples you can copy, and the parts most guides leave out, about cost, reliability and knowing what not to let run on its own.

Both this and the prompt engineering book also come as one 12-in-1 volume, if you would rather have the whole system in a single place.