Plan-and-Solve Prompting

Prompt the model to form an overarching plan before working the problem step by step, without any tool use or think-act-observe loop.

Rather than jumping straight into a ReAct-style think-act-observe loop, plan-and-solve prompting asks the model to first form an overarching plan before executing it: "Let's first understand the problem and devise a plan to solve the problem. Then, let's carry out the plan and solve the problem step-by-step." No tool usage is involved — it is purely a reasoning-quality technique, closer in spirit to zero-shot chain of thought prompting than to ReAct's tool-using loop, decomposing the problem before committing to any intermediate answers.

Because it produces a plan without any tool interaction, it is the prompt-level cousin of hierarchical planning, addressing the same generate-a-roadmap-first idea as a single-turn reasoning prompt rather than an agent architecture. Combining preplanning with ReAct's loop — plan first, then think-act-observe against that plan — is a natural harness pattern worth trying when goals need both structure and external actions, or when neither technique alone is enough.

---

Sources

• Prompt Engineering for LLMs (John Berryman), ch. 8