PM-301e · Module 1

Tool Selection Prompting

4 min read

Tool descriptions govern default tool selection behavior. The task prompt governs whether tool use is activated at all, which tools are appropriate for the task, and whether the model should use tools explicitly or implicitly. Three failure modes: over-calling (using tools when not needed), under-calling (not using tools when needed), and wrong-tool selection.

  1. Explicit Tool Use Instructions State in the system prompt when tools should be used: "Use the search_customers tool whenever a customer name or account is referenced. Do not rely on your training data for customer information — always verify via the tool." Explicit instructions prevent under-calling on tasks where tool use is mandatory.
  2. Implicit Tool Use If the task is clear from context and tool descriptions are precise, the model should select tools without explicit instruction. Implicit tool use is appropriate when: tool descriptions unambiguously specify when to use each tool, and the task prompt provides sufficient context for the model to match task to tool.
  3. Tool Use Priority When a task could be answered from training data or from a tool, specify which should take precedence: "Prefer tool results over your training data. If a tool is available that can answer the question, use it." This prevents the model from answering from stale training data when a live data source is available.

Do This

  • Include tool use policy in the system prompt for agents that must use tools
  • Specify which tool to use when multiple tools are plausible
  • Specify when NOT to use tools: "For general knowledge questions, answer directly without using tools"
  • Test tool selection on edge cases where the model might choose incorrectly

Avoid This

  • Assume tool selection is automatic and correct without testing
  • Leave tool priority (tools vs. training data) unspecified
  • Use a single instruction like "use tools as needed" without specifying when "as needed" applies
  • Over-restrict tool use with excessive "only use this tool when..." conditions that prevent legitimate use