Tool-Trigger Calibration via System Prompt
Tune how eagerly a model reaches for tools by writing graded system-prompt instructions and choosing when to force a call instead.
Three graded instruction variants (conservative, balanced, aggressive), a recommendation for the stated failure mode, and a note on when to force a call at the API level instead.
Role
You are calibrating tool-invocation behaviour for a deployed agent.
Task
For an agent whose purpose is {{agent_purpose}} with tools {{tool_registry}}, currently exhibiting {{current_failure_mode}}, write three system-prompt instruction variants that shift tool eagerness: conservative (model uses judgement), balanced (investigate with tools before answering), aggressive (always consult a tool before responding). For each, list the request types it will mishandle. Then map {{no_tool_cases}} and {{must_tool_cases}} into explicit prompt clauses, and state when {{forcing_mechanism}} should replace prompt steering entirely.
Context
With automatic tool choice the model re-decides every turn, so eagerness is a prompt-level dial; guarantees, however, belong in the request configuration rather than in prose.
Inputs
- {{agent_purpose}}
- {{current_failure_mode}}
- {{tool_registry}}
- {{no_tool_cases}}
- {{must_tool_cases}}
- {{forcing_mechanism}}
Constraints
- Each variant must be a single quotable instruction
- Name the trade-off for each variant
- Do not claim prompt phrasing can guarantee a call
- Keep stable-knowledge and conversational turns tool-free where listed
Output Format
Markdown: three labelled instruction variants with trade-offs, a clause table for must/never cases, and a closing recommendation.
Quality Criteria
- Variants are meaningfully different in strength
- Recommendation addresses the stated failure mode
- Forcing mechanism positioned as configuration, not prompting
- No-tool cases explicitly protected