Skip to main content
Lukan uses a split-caching strategy for efficiency, combining multiple components into the system prompt.

​
Cached Block 1 (Stable)

This block rarely changes and is cached for performance:
  • Base agent instructions (or planner instructions)
  • Task tracking rules
  • Sub-agent instructions
  • Browser/Google Workspace/tool restriction rules
  • WhatsApp formatting and directory restrictions

​
Cached Block 2 (Project-Specific)

This block is specific to the current project:
  • Global memory (~/.config/lukan/MEMORY.md)
  • LUKAN.md project guide
  • Project memory (.lukan/memories/MEMORY.md)

​
Dynamic Block (Per-Call)

This block changes with every request:
  • Environment info (directory, platform, date, time, calendar)
  • Pending reminders
  • Current tasks
  • Active plan content
  • Available skills listing

​
Prompt Generation Flow

1. Load Cached Block 1 (base instructions)
2. Load Cached Block 2 (project context)
3. Add Dynamic Block (current state)
4. Append recent conversation history
5. Send to model

​
Performance Optimizations

  • Lazy loading: Blocks are loaded only when needed
  • Incremental updates: Only changed parts are rebuilt
  • Token budgeting: Automatic compaction at 150k tokens
  • Memory updates: Project memory refreshed at 50k tokens

​
Viewing the Prompt

To see the current system prompt: 
lukan doctor
This outputs configuration and diagnostic info including the system prompt, useful for debugging and understanding what Lukan “sees”.