Teach ChatGPT: the five levels, from one chat to a shared assistant

You teach ChatGPT at five levels, and picking the right one is most of the skill. Level one is the conversation itself, where pasted material and examples shape the answers in that chat only. Level two is custom instructions, which apply to all your future chats. Level three is memory, where ChatGPT keeps specific details you ask it to remember and you can review or delete them. Level four is a project or a custom GPT, which holds instructions plus uploaded reference files so a whole body of knowledge is available every time. Level five is the API, where developers supply documents at the moment of the request or fine tune a model on examples. All five teach your own assistant. None of them change what ChatGPT says to anybody else.

That last line matters more than people expect, and the final section of this page comes back to it. First, the five levels and what each is for.

The five levels compared

Level	What it is	How long it lasts	Best for
1. In the conversation	Paste documents, examples and corrections into the chat	That chat only	One off tasks, and testing what instructions actually work
2. Custom instructions	A standing note about who you are and how you want answers written	All your future chats	Tone, format, role, things you never want repeated
3. Memory	Specific facts it saves when you ask it to remember something	Until you delete it	Your business name, your clients, your preferences, recurring context
4. Project or custom GPT	Instructions plus uploaded reference files, grouped	As long as the project exists	A repeated job with a document set behind it, and sharing a configured assistant with colleagues
5. API, retrieval or fine tuning	Your application supplies the documents, or you train on example pairs	Permanent, in your own system	Products, support bots, anything serving customers at volume

Level 1: teach it inside the conversation

This is where most people should start, because whatever works here is what you then promote to a higher level.

• Give it the source material, pasted in or uploaded, rather than expecting it to know your business.
• Give it three examples of good output and one of bad, and say what makes each one good or bad. Examples teach faster than adjectives.
• Give it the decision rules you use yourself: when to recommend the premium option, which claims legal will not approve, what we never say about competitors.
• Correct it once, properly. Instead of "no, try again", say what was wrong, why, and what to do instead. Then ask it to restate the rule in its own words so you can check it understood.
• Save what worked. At the end of a good session, ask it to write the instruction set that would have produced that output first time. Keep that text. It is your prompt for next time and the draft of your custom instructions.

Level 2: custom instructions

Custom instructions are a standing brief applied to your chats. Keep them short and specific, because a long one dilutes itself.

WHO I AM
I run a two location dental practice in Leeds. I write for nervous patients
and for referring GPs.

HOW TO ANSWER
- British English, short sentences, no exclamation marks.
- Lead with the answer, then the reasoning.
- Ask me one clarifying question if the request is ambiguous.
- Flag anything that would need a clinician to approve before publishing.
- Never invent a statistic, a price or a source. Say what you would need.

Level 3: memory

ChatGPT can retain details across chats. The practical way to use it is to be deliberate: tell it explicitly what to remember, and check what it has saved.

• Say "remember that" followed by the specific fact, rather than hoping it picks the right thing out of a long conversation.
• Store durable facts: your company, your market, your tools, your house style, the names of recurring clients or products.
• Keep temporary context out of memory. A deadline next Tuesday is not worth remembering in March.
• Review the saved memories in settings periodically and delete what is stale. Out of date memories cause confident, wrong answers.
• Use a temporary chat for anything you do not want retained, and remember that memory is per account, so it follows you rather than your team.

Level 4: projects and custom GPTs

When the same job comes round every week and has documents behind it, move up a level.

• A project groups chats with shared instructions and files, so you are not re uploading your brand guide every Monday.
• A custom GPT takes instructions plus knowledge files and can be shared with colleagues, which turns your prompt craft into something a team can use.
• Curate the knowledge files. Five accurate, current documents beat fifty, because contradictory sources produce confident contradictions.
• Date every file in its name and in a header line, and say in the instructions which file wins if two disagree.
• Write the refusals in. Tell it what to do when the answer is not in the files: say so and ask, rather than fill the gap.
• Upload nothing confidential that you would not want a colleague to open, and check your organisation's data policy before uploading client material.

Level 5: the developer options

• Retrieval, where your application finds the relevant passages and sends them with the question. This is the right default for support bots and internal search, because the answer is grounded in a document you control and you can show the source.
• Fine tuning, which teaches a model a format, a style or a classification pattern from many example pairs. It shapes behaviour rather than adding facts, so use retrieval for knowledge and fine tuning for form.
• Evaluation, which is the part teams skip. Write twenty real questions with the answers you expect, and run them after every change. Without that set you are guessing.

A business brief worth pasting in

BUSINESS: name, what we sell, who buys it
CUSTOMERS: the three segments, and what each one cares about
PROOF: the facts we can state, with the source of each
WORDS WE USE: our terms for our products and features
WORDS WE DO NOT USE: banned phrases, competitor names, unapproved claims
TONE: three adjectives and one sentence of example copy
RULES: what always needs human approval before it goes out

Fill that in once, keep it in a file, and paste or upload it at whichever level you are working. It is the single highest return hour you can spend on this.

This page covers teaching ChatGPT about you and your work. For using ChatGPT to teach a subject to students, the prompting approach is different and deserves its own guide.

The Teach ChatGPT questions worth an answer

Does teaching ChatGPT change the model for everyone?

No. Custom instructions, memory, projects and custom GPTs shape your own assistant and the people you share a GPT with. The underlying model is trained separately by the provider, so what you teach stays on your side of the account.

Is uploading a file the same as training?

No, and the difference is useful. An uploaded file is reference material the assistant reads when answering. Training changes a model's behaviour from many examples. For almost every business need, giving it the right documents is the better and cheaper route.

Why does it forget what I told it?

Because the level you used was the conversation. Instructions given in one chat do not carry to the next unless they are in custom instructions, memory or a project. If something matters every time, promote it a level.

How do I stop it inventing facts about my business?

Give it the facts in writing, and add an explicit rule: if the answer is not in the material supplied, say so and ask. Then test it with a question you know is not covered and check that it refuses.

What should I never put in?

Anything you are not permitted to share, including client data covered by a confidentiality agreement, personal data you have no basis to process, and credentials. Check your organisation's policy first, and use a configured workspace if one is available to you.

Teaching your own assistant is not teaching your buyer's

The five levels shape how your chat, your project or your shared assistant answers you. A buyer's assistant has never read your brief. When they ask it about your category, it draws on the public web and its own training, and it replies to them in its own words.

So test what it says without your instructions. Use a fresh chat with no memory and ask about your company as a stranger would. AI Knows Us does the same across assistants for the questions you choose, which shows what a buyer sees.