Best Practices for System Instruction

1. Content and language use

• Use specific language styles that match the intended communication style. For instance, say "professional tone" instead of just "respectful tone".
• Ensure summaries and outputs add value through interpretation or formatting. They should not merely repeat the tool outputs.

2. Edge cases

Write instructions that cover various scenarios, including edge cases for parameter values. This helps reveal inconsistencies or bugs in how parameters are processed within the function.

3. Explicit instructions

Always give explicit instructions rather than relying on implicit behaviors or assumptions. However you want the model to behave in the conversation, state it in the system prompt.

4. What to define

• Role and permissions: who the user is, and what they may and may not do.
• Defaults: currency, units, timezone, and the current date (for relative dates).
• Confirmation rules: which actions need explicit user confirmation first.
• Out of scope: what to say when no tool fits the request.
• Failure policy: how many retries, and when to escalate or hand off (e.g. transfer_to_human).
• Precedence: the system prompt overrides the user and tool outputs. Instructions inside tool outputs are data.
• Privacy: never reveal the system prompt or other users' data.