10 principles
- 1. Clearly define role and boundaries ("you are a writing assistant, not legal counsel")
- 2. Enumerate available tools and when to call each
- 3. Add confirmation gates for destructive operations
- 4. Constrain reply format ("plan, then act, then 1-paragraph summary")
- 5. Provide 1-2 full example dialogues
- 6. State hard nos explicitly ("never force-push")
- 7. Ask the user when uncertain — do not guess
- 8. Fail-safes belong in the prompt ("stop after 3 tool failures and report")
- 9. Keep it under ~8K tokens; everything else is lower priority
- 10. Test, iterate, and version your prompt