The system prompt, the instructions you don't see

When you talk to a chat you think the conversation begins with your first message. It doesn't begin there. Before you write anything, the company has already given the model a text you never get to see, and understanding that this text exists changes how you read everything the chat answers you.

In the previous steps I went through the model's quirks: that it hallucinates, that it varies, that each brand has its character. That character doesn't come only from how it was trained. A good part comes from here, from a hidden script I'm going to show you.

What the model reads before you do

That script is called the system prompt, the system instructions. It's a text the company places before the conversation and the model reads first, before your message. It tells it how to behave: what tone to use, what to call itself, what it may tell and what it must avoid, in what language to answer. It's, so to speak, the house rules.

Remember what we saw at the start of the staircase: the model only works with the text it receives. It doesn't tell apart in some mystical way "what the company says" and "what you say"; it all reaches it as text. What happens is that the maker puts its part first, and that part shapes what the model builds afterward. You write on top of a text that was already there, even if you don't see it.

That's why, when you ask a chat "who are you?" and it answers with a name, an attitude and a few things it says it can't do, almost none of that did the model decide on its own. It was told so in the system prompt, and it repeats it.

The page is never blank

Here's the misunderstanding worth knocking down. It's easy to picture the chat as a blank page, a neutral assistant that reacts only to what you ask of it. It isn't. It arrives configured from the factory, with priorities and limits you didn't choose and don't see.

The image that worked for me is that of a sheet of paper that looks empty but is written in very faint ink. At a glance you see nothing and think you're going to start from scratch. In reality, on that sheet there's already a heap of instructions in place, and you write on top. Your message doesn't land in a void: it lands on a previous text that steers the answer.

This is nothing shady or secret. It's normal configuration, the way a company makes its product behave in a particular way and not just any way. It's only worth knowing that text exists, because it explains many things you'd otherwise pin on the model.

Who has the upper hand

There's one detail that's the one that really matters, and it's who the model listens to more when you and the company say things that don't match. The answer is clear: the company. The system instructions weigh more than yours. Anthropic, the company that makes Claude, describes it as the highest-priority message within the conversation, above whatever the user writes.

This shows a lot. If you ask the chat for something its system prompt forbids, it won't do it no matter how much you insist, because it has an underlying instruction telling it no, and that instruction beats yours. When a chat flatly refuses something, often it isn't the model "deciding" in that instant: it's that invisible text giving the orders.

Knowing this changes how you read a refusal. It stops looking like a whim of the machine and becomes what it is: a house rule applied before you open your mouth. You aren't arguing with the model; you're running into its configuration.

A whole document, not a sentence

What surprised me most was the size. I pictured the system prompt as a couple of lines, something like "be kind and don't say silly things." It's far more. It's a whole document of rules that accompanies every conversation.

In 2026, The Washington Post analysed the system prompts of several well-known chatbots and published the figures: they ranged from around 2,300 words in the shortest to around 27,000 in the longest. Twenty-seven thousand words of underlying instructions, more than many short stories. Inside there are rules of all kinds, from tone ("aim for readable, accessible answers") to very specific limits. The report noted, for instance, that Claude devotes more than two thousand words just to avoiding copyright problems, with a list of how many words it may quote from an article or how many lines of a song.

When you see that volume, you better understand why two chats that underneath use similar models behave so differently. A large part of the difference is in those thousands of words of instructions each company writes in its own way and that you never read.

When the chat dodges something

With this in mind, many of the chat's odd behaviours stop being a mystery. Does the chat dodge a topic and change the conversation? Does it always go on with the same courteous formula? Does it repeat a style bias, like rounding off every answer with a summary? A good part of the explanation is in its factory instructions, not in an opinion of the model.

This helps you not confuse two things that are very easy to mix up: the company's rules and "what the AI thinks." When the chat refuses to give an opinion on something controversial or carefully repeats a lukewarm stance, you aren't seeing the thought of a mind; you're seeing a product decision someone wrote in that hidden text. It's the company speaking through the model's mouth, not the model thinking on its own.

With this I close off what lies beneath the chat: how it works inside, its quirks and, now, that underlying script that goes with it without you seeing it. From here I stop looking at the machine's side and start looking at yours, what you write. Because on top of that prior text, the next factor that most changes the answer is how you talk to it: the prompt.

Definitions

- System prompt (system instructions): the text the company places before your conversation to set how the chat behaves: tone, name, limits, what it can and can't do. You don't see it, but the model reads it first. - System prompt priority: the fact that the model listens more to the company's instructions than to yours. If they clash, the house's win. - Factory instructions: another way to name that hidden script. They come pre-set by the maker; you didn't choose them.

No comments yet

No comments yet. Be the first.