Planning Prompts

Introduction

A planning prompt asks the model to write the plan first (steps, sub-questions, sources) and then execute it. Foundational for agents.

Beginner analogy: A craftsperson sketches before cutting wood. The plan IS the prompt for the next prompts.

Understanding the topic

Core concepts to understand:

• Step 1: 'Write a plan as a numbered list.'
• Step 2: Execute each step (often in separate LLM calls).
• Plan can be revised between steps.
• Backbone of ReAct, Plan-and-Execute, AutoGen frameworks.

Syntax reference

Visual workflow / architecture:

Plan:
  1. Search for X.
  2. Summarise findings.
  3. Identify gaps.
  4. Ask user.
  5. Finalise.

Now execute step 1...

Real-world use

Every AI agent that does multi-step work — booking flights, debugging code, writing reports — starts with a planning prompt.

Best practices

• Cap plan to 5–7 steps; more = more drift.
• Allow re-planning when execution reveals new info.
• Show plan to user for transparency where appropriate.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why generate the plan as a prompt?
• Q2. When should you re-plan?
• Q3. Plan + reflection — combine how?
• Q4. Risks of planning prompts?