System prompt
The standing instruction layer that shapes every response a model gives — role, boundaries, format. Where preferences belong; where information does not.
01What it means
A system prompt is the instruction a model receives on every request: who it is, how to behave, what format to answer in, what never to do. Unlike a user message, it is not part of the conversation — it is the standing policy the conversation happens inside. Every serious AI product carries one; the question is what it contains and why.
The diagnostic for a healthy system prompt is its growth rate. Every rule in one is either a genuine preference — tone, format, boundaries — which belongs there, or a patch for missing information: "the customer's plan is X", "inventory updates arrive hourly", "these three fields are unreliable". The patches are the smell: each one is context masquerading as instruction — information the model should know, written down as behavior it is told to perform. A ten-page system prompt usually means someone is compensating for missing context, not model limitations.
As models reason better from context and intent, prose stopped being the cheapest place to put knowledge. The durable design moves information out of the prompt and into retrieval, memory, tools, and user state — assembled per request — and reserves the prompt for what it is good at: preferences and policy.
02Why it matters for buyers
- —Prompt length is an audit tool: ask an AI vendor how long their system prompt is and why. "Ten pages and growing" signals patched design; "short, because the facts live in context" signals the opposite.
- —The preference-versus-information split predicts what breaks: preferences age well in prompts; information goes stale silently — and stale instructions are how confident wrong answers happen.
- —Red lines belong here: data classes and prohibited actions must sit in the system prompt, where they apply to every turn, not in a user message the agent may never re-read.
03Common questions
What belongs in a system prompt?
Preferences and policy: role, tone, format, boundaries, red lines. Not facts the application can supply as context — those belong in retrieval, memory, tools, or state.
How long should a system prompt be?
As long as its preferences require and no longer. Growth is a symptom, not a virtue: every new rule is either a preference (fine) or missing information wearing a costume (move it to context).
What is the relationship to context engineering?
The system prompt is one input among many; context engineering designs all of them — retrieval, memory, tools, state — and deliberately keeps the prompt for policy.