The Idea
When AI tackles a complex task in one shot, it often loses track of what it's supposed to do. It might skip steps, forget earlier context, or get distracted by a difficult sub-task and lose sight of the big picture.
Plan-and-Execute solves this by separating the two jobs: first, a planner creates the complete step-by-step plan. Then, an executor works through each step one at a time. The planner sees the whole picture. The executor focuses on one step. Neither is overloaded.
Building Blocks
This composition extends:
Plan First Think Step by StepWhere "Plan First" is a single-prompt hint ("plan then answer"), Plan-and-Execute makes it a true two-phase system with distinct planner and executor roles — and optional replanning when things go off track.
The Three Phases
Plan
Create a complete step-by-step plan before doing anything. Think about what needs to happen and in what order.
Execute
Work through each step one at a time. Focus fully on the current step. Use results from earlier steps as context.
Replan
If a step fails or reveals new information, update the plan. Adapt without losing the overall structure.
See It in Action
Task: "Compare the top 3 note-taking apps and recommend one for students."
2. Research features and pricing for each
3. Evaluate pros and cons of each
4. Create a comparison table
5. Analyze which best fits student needs
6. Make final recommendation with reasoning
Research Obsidian: free, Markdown, offline-first, plugin ecosystem...
Research OneNote: free, Microsoft integration, handwriting support...
Why This Works
Trying to plan and execute simultaneously is like driving while also reading the map. Separating the two lets each phase do its job well: the planner thinks broadly about strategy without getting lost in execution details. The executor focuses narrowly on each step without worrying about the big picture.
This also makes it practical to use different AI models for each phase — a powerful model for planning (where strategy matters) and a cheaper, faster model for execution (where the task is clearly defined). You get better results for less cost.
The Composition
Create the complete plan first with a strong planner. Then execute each step with a focused executor. If something goes wrong, replan and continue. Thinking and doing are separate jobs.
How to Apply This
- "First, create a detailed step-by-step plan for this task. Don't execute anything yet — just plan."
- Review the plan, adjust if needed
- "Now execute step 1 of your plan." Then: "Now execute step 2, using the results from step 1."
- If a step reveals something unexpected: "Given what we learned, update the remaining plan."
When to Use This
- • Complex tasks with many steps where AI keeps losing track
- • Research tasks with a clear workflow: search, analyze, compare, recommend
- • When you want to review the plan before any work begins
- • Long tasks where a single prompt would exceed the conversation's focus
- • When you want to use a cheaper model for straightforward execution steps
When to Skip This
- • Highly dynamic tasks — if you can't predict the steps in advance, a reactive approach works better
- • Simple questions — planning overhead isn't worth it for quick answers
- • Tasks needing constant adaptation — if every step changes what the next step should be, consider a more adaptive technique
How It Relates
Plan-and-Execute is the multi-call evolution of Plan First. Where Plan First asks AI to plan and execute in one prompt, Plan-and-Execute makes each phase a separate step with its own context and focus.
It contrasts with reactive approaches like the think-act-observe loop (used in agent systems), which decides what to do one step at a time based on what just happened. Plan-and-Execute is better when you can predict the workflow upfront; reactive approaches are better when you can't.