Plan-and-Solve Prompting
What You'll Learn
You'll learn how to improve LLM problem-solving by explicitly separating planning from execution—first asking the model to devise a strategy, then working through it step-by-step.
The Core Idea
Plan-and-Solve Prompting is a two-phase approach to complex reasoning tasks. Instead of jumping straight into solving a problem (even with Chain-of-Thought reasoning), you first ask the model to create a plan outlining the steps needed, then execute that plan methodically.
Think of it like cooking a complicated recipe: a chef first reads through all the steps, organizes ingredients, and plans the sequence (mise en place) before starting to cook. They don't just start chopping and heating randomly while reading instructions line-by-line.
How It Works
The technique typically follows this structure:
- Planning Phase: Prompt the model to analyze the problem and generate a high-level solution strategy
- Execution Phase: Follow the generated plan step-by-step to reach the final answer
For example, instead of "Solve this word problem: [problem]", you'd prompt:
"First, devise a plan to solve this problem. Then, carry out that plan step-by-step to find the answer."
This separation helps the model:
- Avoid missing critical steps
- Reduce errors from jumping to conclusions
- Better handle multi-step problems requiring different operations
- Create more structured, verifiable reasoning
When to Use It
Plan-and-Solve shines with complex problems requiring multiple distinct operations—especially math word problems, multi-step logical reasoning, or tasks where order matters. It builds on Chain-of-Thought by adding explicit strategic organization before execution.
Key Takeaway: Plan-and-Solve Prompting improves reasoning quality by forcing the LLM to think strategically first (make a plan) before diving into tactical execution (solve step-by-step).