What a custom assistant actually is

Every major assistant now offers a way to save a configuration and give it a name. ChatGPT calls them custom GPTs, Claude has Projects, Gemini has Gems, and the pattern is the same everywhere because the underlying idea is simple.

Simpler, in fact, than most people assume — and the gap between what it is and what it is believed to be costs real money.

Three parts, and that is all

Instructions. A block of text sent with every request. The thing P1 taught you to write, saved so nobody retypes it.

Knowledge. Optionally, some files. The platform does retrieval over them — the machinery from P2, with the choices made for you.

Permissions. Whatever the platform lets it reach: a search tool, your files, a connected service. Often none.

That is the whole object. A custom assistant is a saved configuration.

Nothing is being trained

This is the misconception worth killing first, because three expensive mistakes grow from it.

When you write instructions into a custom assistant, the model does not learn them. They are prepended to every conversation, invisibly, exactly as if you had pasted them yourself and then deleted the paste from view. The model is the same model everyone else is using. Your configuration changes what it is told, not what it knows.

Three consequences follow, and each is something people genuinely get wrong:

It does not improve with use. A custom assistant that gave a poor answer in March gives the same poor answer in October unless somebody edits the instructions. Nothing accumulates. There is no learning curve to wait out.

Your files have not been absorbed. They sit alongside, and are searched when a question seems relevant. If retrieval misses, the answer is as if the file did not exist — which is P2·L4’s entire subject, and it applies here with the ranking hidden from you.

You almost certainly do not need fine-tuning. Fine-tuning does change a model’s weights, costs considerably more, needs a dataset, and is the wrong tool for “I want it to follow our house style”. A configuration handles that. People reach for fine-tuning because the configuration felt too simple to be the answer, and the simplicity is the point.

What the saving actually buys

If it is only a saved prompt, why bother?

Nobody retypes it. The thing P1·L8 said about prompts that survive — this is the mechanism for surviving.

Everyone gets the same one. Eight colleagues with eight slightly different prompts produce eight slightly different standards. One configuration produces one.

It can be corrected centrally. Find a flaw, fix it once, and every future conversation has the fix. That is a genuine shift from pasted prompts, and it is the strongest argument for the feature.

It lowers the floor. A colleague who would never write a good prompt can use a good one, which is often the real reason to build one at all.

What it does not fix

A configuration cannot make a model know something it does not know, and cannot make it reliable at something it is unreliable at. If the underlying model gets your domain wrong, a custom assistant gets your domain wrong in a consistent house style — which is arguably worse, because consistency reads as competence.

Lrnon’s own, as a worked example

Sky is a custom assistant built by hand rather than on a platform, and it has exactly the three parts:

Instructions — a system prompt with five numbered rules, sent on every request
Knowledge — a retrieval index built from the lessons at build time
Permissions — none worth the word. It reads the index and writes an answer. It cannot send, fetch, or read any learner’s record.

Nothing about it is trained. The rules are re-sent every time, and when one of them was found to be insufficient in September the fix was a sentence added to a text file. Each of those properties is a platform feature too; building it by hand only makes them visible.

Try it now (6 minutes)

Open whichever assistant you use and find its saved-configuration feature. Do not build anything yet. Just read the form and name each field as one of the three parts above.

Then answer one question: if this is only a saved prompt, what would I gain by saving it? If the answer is “nothing, I am the only user and I only do this occasionally”, you have learned something useful, and L8 is about exactly that case.

Check your understanding

1. What happens to the instructions you write into a custom assistant? The model is trained on them and remembers them They are sent with every request, like a prompt nobody has to retype They are compiled into a private copy of the model

2. A custom assistant gave a weak answer in March. By October, with heavy use, it will: Have improved, because usage is feedback Give the same weak answer, unless somebody edited the instructions Have improved only if users corrected it in conversation

3. The strongest practical argument for saving a configuration rather than pasting a prompt is: It uses fewer tokens per request A flaw can be corrected once and every later conversation has the fix The platform optimises saved instructions automatically

Recap

A custom assistant is instructions, optional knowledge and whatever permissions the platform grants — a saved configuration and nothing more. Nothing is trained, so it does not improve with use, your files are searched rather than absorbed, and fine-tuning is almost certainly not what you want. What saving buys is one shared standard that can be corrected centrally. What it cannot buy is competence the underlying model lacks.