System prompts: the instructions you never see

A system prompt is a block of text an app places in front of your message before the model ever sees it. It typically says who the assistant is meant to be, what it should refuse, what tools it can call, today's date, and how answers should be formatted. You do not write it, you usually cannot read it, and it is the single biggest reason the same underlying model feels like a different product in two different apps. It is not a control panel or a permission system. It is a prompt, written by someone else, that arrives first.

What the model actually receives

The model does not have compartments. Everything reaching it is one sequence of text with role labels attached: this part is system, this part is user, this part came back from a tool. There is no separate channel and no locked drawer. During training, the model is taught to treat text labeled system as higher priority than text labeled user, which is why those instructions usually win. But "usually wins" is a learned habit, not a wall, and that distinction explains almost everything odd about how assistants behave, including why people can sometimes talk a model out of its instructions. Underneath, it is still predicting the next token from everything in front of it.

That last row is the one people forget. When an assistant reads a web page or a shared document for you, that text lands in the same stream as the instructions, and the model has no dependable way to tell content apart from commands. Words planted in a document saying to disregard earlier instructions are a real attack, known as prompt injection. In a plain chat the damage is limited to a strange answer. It becomes serious when the assistant can act on the world, which is the permission problem covered in what happens when a chatbot starts taking actions.

Why the same model behaves differently in two apps

People often conclude that a model got worse or that one company has a smarter version. Usually four less exciting things are different.

The system and developer prompts differ. One product asks for concise answers with bullet points, another asks for a friendly conversational tone, a third forbids the assistant from discussing anything outside its product. That alone produces answers that feel like different personalities.

The available tools differ. An assistant that can search the web answers a question about last week; one that cannot will fall back on what it learned during training, with all the confidence and none of the currency.

The injected context differs. Apps quietly add things: the current date, your location, the document open beside you, a summary of earlier chats, retrieved company material. Same question, different accompanying material, different answer.

The settings and the tier differ. Randomness settings, response length caps and sometimes a smaller or larger version of the model sit behind the same brand name, which is often what people are actually noticing when a free and a paid interface disagree; what a subscription actually buys you goes through those differences.

Custom instructions are your own system prompt

Most assistants let you save standing instructions, usually split into something about you and something about how you want replies written. Mechanically, this text is stored by the app and added near the front of every new conversation. It is not enforced by the software. It is more text, competing for attention with everything else, which explains both its strengths and its failures.

It works well for durable things. Where you are and what units you want. The vocabulary you already know, so it stops explaining basics. A default output shape, such as no restatement of the question and no closing summary. A hard preference like plain text rather than headings. These hold up because they apply to every request and never contradict anything you type later.

It works badly for anything aspirational. "Always be accurate" and "never make things up" have no lever behind them; there is no honesty dial, and a model that could tell when it was wrong would not be wrong. "Always be brief" backfires the first time you want depth, because the standing instruction and the specific request now disagree, and the ambiguity produces something in between. Keep the list short for a further reason: every word sits in the context of every conversation you have, gets processed on every turn, and dilutes everything else in there.

Custom instructions are also not the same thing as memory. Memory is written by the assistant about you over time and pulled back in when it seems relevant, so it changes without you touching it, which is a different set of trade offs covered in what the assistant keeps about you. Both are stored on someone else's system, so treat the settings box the way you would treat any other field you type personal details into, with what a chatbot does with personal data in mind.

Why instructions fade in a long chat

You ask for short answers at turn two. By turn thirty you are getting essays again. Three things are happening at once.

The first is dilution. Your instruction was one sentence in a conversation that is now thousands of words long, and recent text exerts more pull than old text. The second is compaction: once a conversation approaches the limits of the context window, earlier turns get summarized or dropped, and precise wording is exactly what a summary loses. The third is contradiction by example. You asked for brevity, then asked five questions that needed detailed answers and did not complain, so the recent pattern of the conversation now argues against the old instruction.

The fix is unglamorous and works: restate the constraint in the message where it matters, rather than trusting something you said twenty turns ago. Start a new chat when the task changes, since a fresh conversation is the cheapest way to clear accumulated drift. And keep style rules attached to requests rather than to sessions.

What a persona prompt can and cannot do

Telling a model it is a senior tax attorney does change the output. It shifts vocabulary, structure and what the answer treats as relevant, and it can be genuinely useful for that reason. It does not add knowledge the model never had, it does not grant access to anything, and it does not improve accuracy. The uncomfortable part is that it often increases confidence, so the answer sounds more authoritative while being exactly as likely to contain an invented detail as before.

In practice the useful half of a persona line is the audience and the framing, not the job title. "Explain this to someone who has never filed a return, using an example with real numbers" changes the answer far more than "you are an expert", because it describes what the output should look like rather than who is supposedly writing it. That is the same principle as writing a brief rather than a wish.

The custom instructions worth writing today

Open the settings and write five lines, no more. One line of durable context about your work or where you are. One line on what you already know, so it stops over explaining. One line on default shape, such as answer first, no restatement of the question, no closing summary. One line on format, such as prose by default and lists only when the content is genuinely a list; the specifics of that are in how to get the exact format you asked for. One line on what to do when it is unsure, such as saying so and naming what would resolve it, which is a request for a visible signal rather than a promise of accuracy. Then leave everything task specific out, because that belongs in the message, not the settings. Reread the list in a month and delete whatever you have been silently overriding in every chat, since that line is costing you attention on every single turn.