System prompt

The system prompt is the block of instructions the harness places at the top of every model provider request, ahead of your messages. It tells the model what role it is playing, which tools exist and how to call them, what the house rules are and, in coding agents, a good deal about how to work: when to ask, how to format edits, what to check before declaring a task done. You rarely see it, but the model sees it first on every single turn.

Because it is resent each time and sits at the front of the context, the system prompt has outsized influence, and it explains a lot of behaviour that looks like the model's personality. Claude Code asking before a destructive command, Codex writing a summary at the end of a task, Cursor preferring small diffs: those are harness instructions. When two products on the same model behave differently, the system prompt is usually the first place the difference lives.

You do not edit the vendor's system prompt, but you extend it. Project instruction files such as AGENTS.md and CLAUDE.md are appended into the same request, which is why a rule written there feels 'built in' while a rule typed in chat fades as the session grows. Keep those files short and specific; they cost input tokens on every request, and a bloated one dilutes its own rules.