Forced Reasoning Before Tools

Disable tool choice for a turn while still declaring tools so the model must plan in language before it is allowed to call functions.

When a task jumps into function calling without planning, force a reasoning turn first: set tool_choice to "none" (or the provider equivalent) while still including the tool schemas in the request. The model sees what tools it will have next turn and must think in prose — often with a chain of thought or plan-and-solve cue — before a later turn enables calls.

This is a harness control for task quality inside templated prompt tasks and ReAct-style loops: separate "think about the inventory" from "invoke." Some models (for example Claude Opus) reason by default; others need the explicit nudge. Pair with clearer prompts before escalating to Reflexion retries. It is a narrower, single-turn version of the same thinking/acting split behind plan-validate-execute — a request-parameter toggle rather than an architectural one.

---

Sources

• Prompt Engineering for LLMs (Berryman), ch. 9