Prompts

Prompts control how the agent behaves at a fundamental level — what personality it has, what rules it follows, how it responds. You can view, edit, and manage prompts from the dashboard.

Open Agent Capabilities → Prompts. You can:

• View the current system prompt for any agent
• Edit prompts inline
• Save prompts for reuse across sessions
• Switch between saved prompts

The default system prompt lives at agents/prompt.md in the project directory. Custom agents reference their own prompt file in their agent config JSON.

The system prompt is injected at the start of every session. It defines:

• Identity — who the agent is and how it introduces itself
• Behavioral rules — what it should and shouldn't do
• Output format — how it structures responses
• Tool usage patterns — when to use tools vs. respond directly
• Domain expertise — background knowledge relevant to the work

Use prompts for permanent agent-level behavior. Use steering for project rules. Use skills for knowledge that's only relevant sometimes.

Each agent config can reference a different prompt file:

{
  "name": "reviewer",
  "prompt": "agents/reviewer-prompt.md",
  "model": "auto"
}

When you create a custom agent from the Agents panel, you can write its prompt inline or reference a file.

Keep prompts focused. Very long prompts consume context budget that could be used for conversation history, memory, and skills. A good system prompt is typically 200–500 words — enough to establish behavior without crowding out working context.