Build a custom GPT in one paste.

What you need before you start

A paid ChatGPT plan. Creating a GPT requires one. That is the only hard requirement. No code, no API key, no extension. Budget about 20 minutes for the first one.

You also need one job. "Marketing assistant" is a job title. "Turns a bullet list of product changes into a release note in my house style" is a job. Build the second kind and you will keep opening it.

How to create a custom GPT in six steps

The builder's labels change, so I describe each step by what it does.

1. Open the builder. Go to the GPTs area of ChatGPT and choose to create a new one. If the builder offers a chat that writes the configuration for you, skip it and go to the manual configuration panel.
2. Name it and describe it. Put the job in the name, since that is what you will search for in three months. The description is one line for whoever finds it.
3. Write the instructions. This is the field that matters. The next two sections cover what goes in it.
4. Add knowledge files. A style guide, a pricing sheet, ten past examples of the output you want. Only attach files that agree with the instructions.
5. Set conversation starters. These are the buttons a person sees first. Make each one a complete request the GPT was built for, so the first message already carries the inputs it needs.
6. Publish. Keep it to yourself, share by link, or list it publicly. Test it in a fresh chat before you send the link to anyone.

What goes in a good custom GPT instruction

Every one of the 100+ mega-instructions in my toolkit is built from the same four parts.

The role. Who the GPT is, who it works for, and where its job ends. Two or three sentences. The boundary is the part people skip, and it is what keeps a support GPT from handing out legal advice.

The context it should ask you for. List the inputs the GPT needs to do the job well, and tell it to ask for any that are missing before it writes a word. This one rule removes most bad first drafts, because the GPT stops guessing at what you meant.

The frameworks it works from. The method, checklist or structure it applies. A pricing GPT works from your tiers and discount rules. A post GPT works from a hook, one story, one point, one question. Name the method and the output stops wandering.

The exact output format. Headings, length, what is always included, what is never included. Write it as a template with placeholders if you can. The GPT will match a template more closely than any adjective you could add.

Written this way, the GPT behaves the same on its hundredth chat as on its first. And if you would rather skip the builder, the same four-part instruction pasted at the top of a normal ChatGPT conversation does the job for that session.

Three example instruction skeletons

Copy one, fill the brackets, paste it into the instructions field. Each is a skeleton. A full mega-instruction runs several times longer and carries the frameworks in detail.

A customer support GPT:

Role: You are the support assistant for [company], maker of [product].
Answer only questions about [product]; for anything else, point the
customer to [contact channel].
Before answering, ask for: their plan, what they tried, and any error text.
Work from: the attached help docs and refund policy. Never promise a
refund; explain the policy and hand off.
Output: one greeting line, numbered steps, one follow-up question.
Under 150 words.

A LinkedIn post GPT:

Role: You write LinkedIn posts for [name], a [role] who posts about
[topic] in the first person.
Before writing, ask for: the one idea, a real example, and who it is for.
Work from: the attached voice notes and past posts. Short lines. No
hashtags, no emojis.
Output: three post options, each with a different first line, each under
180 words. End with one plain question.

A meeting notes GPT:

Role: You turn raw meeting transcripts or notes into a summary for
people who were absent.
Before writing, ask for: the transcript, the attendees, and what the
meeting was meant to decide.
Work from: keep decisions, owners and open questions. Drop small talk.
Output: Decisions (bullets), Actions (owner, task, date), Open questions,
Next step. Under 250 words. Flag anything unclear rather than guessing.

The meeting notes one is the fastest to prove: run last week's transcript through it and compare the result with the summary a person wrote.

Mistakes that make a custom GPT drift

Drift is when the GPT gives you a sharp answer on Monday and a generic one on Thursday for the same kind of request. Four causes come up again and again.

Instructions that are too short. Two sentences of role and nothing else leaves the model to fill in the method and the format itself, and it fills them differently each time.

No output format. Without a template the GPT chooses a shape per chat. Give it headings and a length and the shape stops changing.

No "ask me first" rule. A GPT that never asks will guess at the missing inputs. Tell it which inputs it needs and to ask before it writes.

Knowledge files that contradict the instructions. Old pricing in a PDF, new pricing in the instructions. The GPT picks one at random. Keep the files current or take them out.

Custom GPT vs custom instructions vs a saved prompt

	Custom GPT	Custom instructions	Saved prompt
What it is	A separate ChatGPT with its own instructions and files	Settings that apply to all of your chats	Text you paste at the start of a chat
Scope	One job, done the same way every time	Your general preferences: tone, format, facts about you	Whatever that one prompt does
Shareable	Yes, by link or listing	Tied to your account	Yes, as text
Knowledge files	Yes	No	No
Best for	A repeatable job, or one you hand to clients or a team	How you want every answer to read	One-off tasks and experiments

Use custom instructions for how you want every answer to read. I wrote a full walkthrough in how to use custom instructions for ChatGPT. Use a saved prompt for anything you run a handful of times; the ChatGPT prompt library has ready ones by job, and the prompt generator builds one for a job the library does not cover. Build a GPT when the job repeats and the setup is identical every time.