System prompt
Instructions given to the model outside the conversation. They apply to every turn, not to one message.
A system prompt holds standing rules: who the model plays, what it always or never does, what shape the output takes. It travels separately from user messages. Models are trained to weight it more and to keep following it as the chat goes on.
Here is a quick test for whether something belongs in it. Would you otherwise paste it into every message? Role, tone, output format, domain limits and refusal rules pass. The task of the moment does not.
Products expose it under different names. ChatGPT calls it custom instructions, the API has a system parameter, and a repository can carry a CLAUDE.md or AGENTS.md file. It is one mechanism with four labels.
Tagged
Where this comes up
Prompts, configs and tutorials in the library that touch this term.
No-padding everyday default
The one to set if you set only one. Answer first, nothing before it, nothing after it, and no offer of further help.
Tested on Claude Sonnet · May 2026
- GuideUniversal
Moving your rules into a config file
How to stop repeating the same instructions in every message, and what belongs in a system prompt against what belongs in the task.
- GuideUniversal
Turn a messy task into a prompt in five minutes
Most weak prompts come from a task nobody fully described. A five-step routine to fix that before you type, with a worked example.
- GuideUniversal
Summaries you can trust
A summary is a claim about a document, so it can be wrong in the same ways as any claim. Four habits keep it honest.
- GuideUniversal
Briefing a coding assistant: the first message
A coding assistant works from what you tell it and what it can see. Here is what to put in the first message so you spend less time correcting it.
- GuideUniversal
How to check what an AI tells you
A few checks that catch most wrong answers: which claims to verify, how to ask for sources, and why a confident tone tells you nothing.