System prompt
The system prompt is the fixed set of instructions a model receives before every conversation: who it is, what it can and cannot do, what tone it answers in and what to do when it does not know something. The user never sees it and it is sent on every request.
What is system prompt?
The system prompt is the fixed set of instructions a model receives before every conversation: who it is, what it can and cannot do, what tone it answers in and what to do when it does not know something. The user never sees it and it is sent on every request.
Also: system instructions · system message
Why it matters
What changes in a SaaS
It is the cheapest thing to adjust and what changes an agent behaviour most. Before considering training a model or switching provider, there is nearly always room here: most strange answers come from ambiguous instructions, not from the model.
System prompt in detail
What it has to carry
What the product is, who it serves, what information it may use, what it must never do — invent figures, promise deadlines — and what to do when a fact is missing. That last point is what separates a careful agent from a dangerous one.
Say what to do, not only what not to do
"Do not invent prices" works worse than "if you are asked a price and it is not in the context, say you will confirm it and hand the conversation to a person". A prohibition leaves the model with no way out; an alternative gives it one.
It is code, even though it looks like text
It changes behaviour in production, so it deserves version control, review and a way to check that a change does not break what was working. Editing it by hand in a panel with no history is like deploying without a repository.
How Intake handles it
Questions about system prompt
How long should a system prompt be?
As long as it needs to cover the cases that matter, and not one line more: it takes up context window on every request and is paid for every time. If it grows without stopping, it usually means somebody is patching case by case instead of fixing the underlying instruction.
Can the user see the system prompt?
They should not, but it is worth writing it as if they could. Attempts to extract it are routine, and a prompt that would say nothing embarrassing if it leaked is a better-written prompt.
Related terms
A term on its own is only half understood. These come up in the same conversation.
LLM — Language model
A language model is a system trained on enormous amounts of text that, given a fragment, predicts how it continues. Writing, summarising, translating and holding a conversation all come out of that one simple capability: they are all ways of continuing a text plausibly.
Context window
The context window is everything a language model can have in front of it at once to produce an answer: the system instruction, the conversation history, the passages retrieved from documentation and the data looked up. It is measured in tokens and it has a ceiling.
Guardrails
Guardrails are the restrictions that stop an AI system doing or saying certain things: which sources it may take data from, which actions it may run, which topics it does not touch and what requires confirmation from a person. They are applied outside the model rather than by trusting it to obey.
Evals — Evaluation
An evaluation is a set of test cases used to check, repeatably, whether an AI system answers as it should. Each case carries an input and a success criterion, and the whole set is re-run every time something changes: the prompt, the model or the documentation.
Hallucination
A hallucination is an answer from a language model that sounds right and is not: an invented figure, a feature that does not exist, a plausible number nobody checked. It is not a bug, it is a direct consequence of how a model works — it produces the most probable continuation, not the verified one.
MCP — Model Context Protocol
The Model Context Protocol is an open standard for connecting AI systems to external data sources and tools. It defines a common way to expose what can be read and what can be run, so one connection serves different assistants instead of writing a bespoke integration for each.
