Module 5 of 5 · Beginner · 6 min read
Prompting reasoning models
Opus 5.5, GPT-6 and Gemini 4 think before they answer. That changes how you should prompt them.
- 1Describe the goal and constraints, not the steps to get there.
- 2Use effort or reasoning settings for hard problems.
- 3Ask for a self-check against your criteria.
01What changed
Reasoning models spend time deliberating internally before they reply, exploring approaches and checking their work. For math, code, analysis and planning, this dramatically improves accuracy.
The old trick of writing "think step by step" and scripting each step can now get in the way, forcing the model to follow your plan instead of a better one it would have found.
02How to prompt them
Give the destination, not the route. Be precise about what a correct, complete answer looks like and let the model figure out how to get there.
- State constraints and success criteria explicitly.
- Raise the effort setting for hard problems; lower it for quick tasks to save time and money.
- Ask it to verify the answer against the criteria before responding.
- Provide all relevant material up front; reasoning cannot invent missing facts.
03When not to use one
Reasoning costs time and tokens. For simple rewrites, classification at scale or quick lookups, a fast model at low effort is cheaper and just as good.
Match the model and effort to the difficulty of the task; the Model Navigator path covers this in depth.
Finished reading?
Mark it complete to track your progress, then test yourself with the Prompting Fundamentals quiz.