System prompt

A system prompt is the set of instructions an AI model receives before the conversation starts. It tells the model who it is working for, what it should and shouldn’t do, how to sound and what background to keep in mind. The user usually never sees it, but it shapes every answer.

Why it matters

The same model behaves very differently depending on its system prompt. One version is a terse code reviewer, another a patient support agent that only discusses your product. For most AI products, the system prompt is a large part of what you are actually buying.

For agents, the system prompt also carries the rules of the job: which tools exist, when to ask before acting, and how to report back. Files like AGENTS.md are a way of adding project-specific instructions to that standing context without editing the harness.

Using it well

Write it like a brief to a capable new hire. Say what good work looks like and why, give an example or two, and state the few hard rules plainly. Long lists of “never do X” tend to backfire, because the model spends attention on edge cases instead of the job.

Keep it stable. A system prompt that stays the same between requests can be cached, which makes each call cheaper and faster (see prompt caching). Put things that change, like today’s task or the user’s files, after it.

What it can’t do

A system prompt is not a security boundary. A determined user or a hostile document can sometimes talk the model out of its instructions, and people regularly get products to reveal their system prompts. Don’t put secrets in it, and don’t rely on it alone to stop an agent from doing something harmful. Enforce the rules that matter in the software around the model, where the model can’t argue with them.

Bring your agent. Bring your team.