Pattern 4 of 10
Read first, plan second, act last
On anything complex, jumping straight to the answer invites rework. Asking the model to read, then plan, then act produces steadier results.
Ask for analysis and a plan before the final output, especially on complex work.
This pattern separates a task into phases: first understand the material, then propose a plan, then carry it out. You can do it in one prompt ("first summarize the file, then propose an approach, then implement") or across turns.
Why it works
When the model commits to output before understanding the problem, it bakes in early misunderstandings. A short planning phase surfaces those before they become a full, wrong answer that you have to unwind.
How to do it
Ask for the plan as a separate step, and review it before letting the model proceed.
Example you can copy
I need to refactor a messy module. Work in three phases and stop after each:
Phase 1 - Read: summarize what this code does and its main problems.
Phase 2 - Plan: propose a refactor approach as a short numbered plan.
Phase 3 - Act: only after I approve the plan, implement it.
Start with Phase 1.When it helps
- Complex, multi-step tasks where a wrong start is expensive
- Working with unfamiliar code, data, or documents
- Agentic tools that otherwise run many steps off a vague instruction
When to skip it
- Simple tasks where planning is slower than just doing it
- On reasoning models, do not add generic "think step by step"; ask for real phases instead
Common mistakes
- Letting the model implement before you review its plan
- Confusing this with generic chain-of-thought on reasoning models
- Making the plan so detailed it becomes the work itself
Frequently asked
Should I tell the AI to think step by step?
On modern reasoning models, a generic "think step by step" is often redundant or can hurt output. Instead, ask for concrete phases such as read, then plan, then act, and review the plan before the model proceeds.