Яндекс.Метрика

System prompt

System prompt explained

Baseline instructions an AI agent gets before any user request: its role, rules of behavior, available tools and limits

A system prompt is the fixed part of a request to a language model, written by the agent's developer rather than the user. It defines who the agent is, how it responds and what it won't do. Users usually don't see it: the model gets the system prompt first, then the user's messages.

A system prompt usually covers:

  • The agent's role and tasks
  • Rules of behavior and tone
  • A description of the available tools
  • Limits: what the agent must not do
  • The response format

Keep the system prompt short and specific, since it takes up the context window in every request. Detailed instructions for specific use cases go into skills that load only when needed.

In Pachca's breakdown of its own agent architecture, the system prompt is built from baseline rules, a list of tools and skills: Markdown files with instructions for working with the tracker, the repo and other services. To change how the agent behaves, you usually just edit a skill's text, not the code.

Learn more in the Help center

Read the article →

Related terms