System prompt
Standing instructions an AI reads before the conversation starts, shaping every reply: its role, tone, rules, and what it knows about you. The app maker writes one you never see. You can add your own on top. In ChatGPT and Claude it appears as custom instructions or project instructions, and one clear sentence changes the output.
In plain English
The system prompt is the briefing before the meeting. Who you are, how you talk, what you never do. The model reads it first and carries it into every answer. You never see the one the app wrote. You can write your own on top.
In the real world
- “Custom instructions” in ChatGPT or Claude settings.
- The instructions box in a Claude project, which applies to every chat inside it.2
- The reason a support bot refuses to talk about anything but returns.
When to use it
- You correct the same thing every chat. Tone, length, “don’t use em dashes.”
- You want the model to play a role. Anthropic: one sentence of role setting changes the output.1
A starter you can paste: “I run a small accounting firm. Write for busy business owners, plain English, short sentences, no exclamation marks, no em dashes. When I paste a draft, tighten it; don’t add new points.”
When not to use it
- One-off tasks. A standing instruction for a one-time job is clutter.
- Facts that change weekly. Put those in the chat, not the system prompt.
Good, bad, ugly
- Good: write it once, stop repeating yourself.
- Bad: write it badly once, get the same wrong thing forever.
- Ugly: a system prompt from six months ago quietly steering today’s work, and nobody remembers it exists.
Key facts
- In chat apps
- "Custom instructions" in settings, or the instructions field of a project
- Who writes it
- The app maker writes one you never see. Yours goes on top.
- How long it lasts
- Every message in that chat or project, until you change it
Common questions
Is the system prompt the same as custom instructions?
For practical purposes, yes. Custom instructions and project instructions are the consumer-app version of a system prompt.
What should I put in mine?
Who you are, who you write for, the tone you want, and the things you always have to correct. Keep it to a page.
Can the model ignore it?
Sometimes, especially in long conversations or when a later message contradicts it. Keep instructions clear and don't bury them.
References
- Prompting best practicesAnthropic, Claude Platform DocsView source
- What are projects?Anthropic, Claude Help CenterView source