Skip to main content

System prompt

The system prompt defines Carina's identity, capabilities, safety rules, and workspace-specific instructions for every turn.

Composition layers​

Prompts are assembled from multiple sources (later layers override earlier where allowed):

  1. Platform base - Core behaviour and tool honesty (agents/system-prompt.ts)
  2. Profile personality - Tone and role from active profile
  3. Skill packs - Domain instructions injected when packs are active
  4. Standing orders - User-defined persistent directives
  5. Memory snippets - Retrieved episodic or workspace memories
  6. Channel overlay - Surface capability contract (web vs Telegram vs CLI)

Surface honesty​

Carina adjusts claims about memory, tools, and Scout based on the active surface (product/surface-capabilities.ts). Cloud BYOK mode may omit episodic memory unless core proxy is enabled.

Editing​

  • Self-host: edit profile files under ~/.carina/profiles/ or use carina personality
  • Cloud: workspace settings and enabled skill packs

Avoid putting secrets in prompts; use env and JIT credential tools instead.

Source references​

  • agents/system-prompt.ts
  • profiles/personality.ts
  • agents/standing-orders.ts (when enabled)