One sentence at the top of a prompt changes everything that follows: tell the model who it is. "You are a senior Linux administrator." "You are a patient math tutor." "You are a security reviewer." This is called setting a role, or a persona, and it steers both what the model leans on and how it says it.
It works because of how the model was built. It learned from an enormous mix of writing - casual forum threads, dense technical manuals, tutorials, code. A role points it at the slice that matches. Tell it it's a senior administrator and it pulls from the manual side of what it saw, not the desktop-help side.
Ask a bare question:
claude -p "How do I find what is using disk space?"
You get a general answer, maybe aimed at someone on a laptop. Now lead with a role:
claude -p "You are a senior Linux server administrator. How do I find what is using disk space?"
The reply shifts toward server tools like du, df, and ncdu, assumes you're
in a terminal, and drops the hand-holding. Same question, sharper answer, because
you pointed the model at the right expertise before asking.
Back in the prompting topics you saw the system message - the instruction that
sits above the conversation and sets the ground rules. A role belongs there,
separate from the user's question. When you type a single prompt to claude,
you just lead with the role sentence and let the question follow. Either way the
idea is the same: set who the model is before you ask what you want.
Keep the role honest and specific. "You are a senior Linux administrator" is useful because it maps to real expertise. "You are the world's greatest genius who never makes mistakes" is not - there's no such body of writing to lean on, so it changes the bragging, not the accuracy.
Next: a role sets who the model is. Tone and audience set how it explains.
Your lab setup
This VM comes preconfigured with your AI keys - just type claude to get started. Your key and URL are already set as $OPENAI_BASE_URL and $OPENAI_API_KEY.