Is CLAUDE.md a system prompt?
Not exactly. Claude Code has its own system prompt. CLAUDE.md is loaded into the context as project instructions alongside it, so it works similarly, but the vendor system prompt still applies.
The standing instructions an application gives a model before the conversation, setting its role, rules and format.
A system prompt is a block of instructions that an application sends to an AI model ahead of the user messages, setting the ground rules for the whole conversation: what role to play, what to do and avoid, what format to use and what tools are available. Users of chat apps usually never see it. Developers write it when building on an API. In coding agents, the tool vendor writes a system prompt, and your project files, such as AGENTS.md or CLAUDE.md, are added to the context alongside it so your rules apply in every session.
The system prompt shapes behavior more than any single message, because it applies throughout. If you build with an API, it is your main lever. If you use a coding agent, you cannot edit the vendor system prompt, but your instructions file plays a similar role for your project, which is why a short, specific one has such a large effect.
A simple system prompt for an internal assistant:
You answer questions about the Acme billing API.
Only use the documentation provided below.
If the answer is not in it, say you do not know.
Reply in under 150 words.For a repository, the equivalent is an AGENTS.md file listing commands and rules. It does not replace the agent system prompt; it is added next to it in every session.
A system prompt is not a security boundary. Users and injected content can sometimes get a model to ignore or reveal it, so never put secrets in it, and enforce real restrictions in code.
All 28 terms are in the AI glossary. For models side by side, see frontier vs open-weight.
Not exactly. Claude Code has its own system prompt. CLAUDE.md is loaded into the context as project instructions alongside it, so it works similarly, but the vendor system prompt still applies.
In most apps they cannot see it directly, but determined users can sometimes get a model to repeat it. Assume a system prompt may leak and keep anything sensitive out of it.
As long as it needs to be and no longer. Clear rules, the required format and key context usually fit in a few hundred words. Long prompts cost tokens on every request.