The system prompt is one of the most important and least understood parts of working with LLMs in production. Here's a clear definition and why getting it right matters more than which model you pick.
A system prompt is the set of instructions given to an LLM at the start of every conversation, shaping how it behaves throughout. Think of it as the role, voice, format, and constraints — the standing orders the model follows on every turn.
When you use Claude or ChatGPT in a product or Project, there are usually two prompts at play: the system prompt (set once, applies always) and the user prompt (what you type each turn).
The system prompt establishes context. Example: "You are a B2B proposal writer at [Company]. Draft proposals matching the voice of the examples in your knowledge base. Output in markdown with sections: Executive Summary, Problem, Approach, Timeline, Investment."
Without a system prompt, you get generic-assistant output. With a well-crafted system prompt loaded with context and examples, you get output that sounds like it came from a senior member of your team.
The difference between mediocre and excellent AI workflows is usually in the system prompt, not in the underlying model.
See the full guide: How to write a Claude system prompt.
In Claude Projects, you set it under "Instructions." It applies to all conversations in that Project.
In Custom GPTs, you set it under "Instructions." It applies to all conversations with that GPT.
Typically 100-600 words. Long enough to specify role, voice, format, and constraints. Short enough that the model attends to all of it.
Not in most consumer products. You usually need to start a new conversation. Developers using APIs can modify per-turn.
Usually not. It's behind the scenes. But in some platforms, sophisticated users can extract or infer parts of it.