Make yours

Nobody trained anything

You paste in some of your writing, ask the model to learn your style, and for a while it works. Then it drifts back. Every guide to this uses the word train, and not one of them tells you that nothing was trained.

That is not a quibble about vocabulary. It is the reason the effect wears off, and once you know what is actually happening the fix becomes obvious and boring.

This sits under the AI writing tells. The short version: put it in a file and hand the file over every time.

Does anything actually get trained?

No. OpenAI's own developer documentation states that each text generation request is independent and stateless, and that multi-turn conversation is implemented by sending the previous messages again as parameters on the next request.

So there is no learning between your messages. There is re-reading. The model appears to remember because the transcript is handed back to it each time, and that is a different mechanism with different failure modes.

The documentation is explicit about the cost, too: even when using the API's own conversation-chaining feature, all previous input tokens are re-sent and charged again on every turn. You are paying to re-supply the context, not to retain it.

Is there a route that does change the model?

One, and it is not available to you in the way the guides imply. Fine-tuning is the only documented process that updates a model's weights, and OpenAI describes the result as a customised model, a separate artefact, rather than a modified ChatGPT that then behaves differently in your chats.

It also needs a real dataset. The documented minimum is ten examples, with improvements reported from fifty to a hundred, and more recommended beyond that. Three pasted paragraphs is not a small version of this. It is a different thing entirely.

And the route is closing. OpenAI's documentation states it is winding the fine-tuning platform down, and that it is no longer accessible to new users. Whatever the guides mean by train, this is not it and will not be.

What does the paste-three-samples method actually produce?

Few-shot prompting, which OpenAI defines as steering a model by including a handful of input and output examples in the prompt, explicitly as an alternative to fine-tuning rather than a version of it.

It works, within its limits, and it is worth knowing what those are. The examples occupy the context window on every request. They compete with everything else in there. And when the conversation ends they are gone, because nothing was retained.

The common next step, asking the model to describe the style it sees in your samples, is weaker still. It converts your writing into adjectives, and adjectives are the least actionable thing you can give a model. Why adjectives do not survive the trip covers that at length.

Why does it wear off after a few turns?

Because the instruction is competing rather than persisting. Each turn re-sends everything, your style examples sit further back among more recent and more specific material, and the model weights what is near.

The documentation notes the related trap: a large accumulated context is not free, and everything in the window is shared between what you sent and what the model produces.

This is also why the effect returns in a fresh chat and then decays again. Nothing degraded. The context was simply supplied again from scratch.

What about custom instructions and memory?

Better, because they persist across chats instead of dying with one. ChatGPT's own product documentation describes custom instructions as being for preferences you want followed across chats, such as your preferred response style.

Note the word preferences. It is a stored setting that gets supplied to the model, not a change to the model. The mechanism is the same one as before, with the re-supplying automated.

That is exactly why writing your rules down as a file is the durable version of this. Each surface has its own documented size limit and its own place to put it, and its own position in the order of precedence.

So what should you do instead?

One more thing worth saying on a page about pasting your writing somewhere. The method here requires pasting nothing: the questions are answered in your browser and your answers are never transmitted.

Where these come from

Every claim above is quoted from one of these, and each was read on the date beside it. If one of them has changed since, the page is wrong and we would like to know.

Nothing remembers, so the fix is to hand it over every time: make the briefing once and reuse it everywhere.