Not Teaching AI to Think
Reasoning models work by tackling the problem themselves before generating an answer. I understand why existing step-by-step instructions or excessive examples can actually be counterproductive, and I grasp the approach of leaving only the goal and output criteria.
Prompting for Reasoning Models
Until now, writing good prompts meant guiding the model to think better.
Asking it to think step-by-step, showing examples, breaking down solutions, and instructing what order to judge things before reaching an answer. Many prompting techniques started from creating a path so the model wouldn't jump straight to the answer.
But reasoning models change this premise slightly.
These models are designed not simply to chain sentences quickly, but to hold onto problems longer and examine conditions more carefully. So the instructions we're used to using don't always help.
If you try to teach a model that's already designed to think by saying "think like this," it can actually get in the way.
Thinking Before Creating Answers
General conversational models are strong at quickly producing natural answers that match user input. For tasks with clear answer directions—like short summaries, sentence polishing, or simple explanations—this approach is sufficient.
Reasoning models work a bit differently. Rather than rushing complex problems to conclusions, they're designed to compare conditions, trace possible paths, and hold onto problems longer before creating answers. For example, models like Claude's extended thinking or OpenAI's o-series models are designed to use more computation in their internal reasoning process.
What matters is not peering into the model's internal thoughts. It's understanding that this is a tool built to use more computation on complex problems.
A reasoning model is less like a model that writes longer answers, and more like a model designed to grip harder problems longer.
When Your Old Methods Get in the Way
With general models, saying "think step-by-step" can sometimes help. It makes the model create intermediate steps instead of jumping straight to conclusions.
But with reasoning models, the same approach doesn't always produce good results. If you strongly specify an arbitrary thinking order to a model already designed to break down problems and examine conditions, you can reduce the room for the model to choose a better path.
The same goes for examples. Good examples show the standard for an answer. But overly detailed examples can lock the model into a specific solution method. Depending on the problem, the model might find a better approach on its own, but an example provided by the user can narrow that possibility.
The problem isn't that CoT or examples are bad. The problem is using the same habits when the tool has changed.
Give Goals and Step Back
For reasoning models, sometimes saying less is better.
Instead of specifying the process in detail, provide clear goals, constraints, and output formats. Just tell it what needs to be achieved, what standards to maintain, and what form the answer should take—leave the intermediate steps to the model.
# Specifying the process
Break down this problem into steps 1, 2, and 3.
First check A, then compare B, and finally conclude with C.
# Giving only goals and criteria
Solve the problem below.
Goals:
- Find possible causes in order of priority.
- Separate certain evidence from speculation.
- Suggest solutions that can be fixed with the smallest modification first.
Output format:
1. Most likely cause
2. Evidence to verify
3. Minimum fix
4. Remaining uncertainty
The second prompt doesn't teach the model in detail how to think. Instead, it shows what counts as a good answer. It leaves the process to the model and fixes only the criteria.
When It Really Shines
Reasoning models aren't a necessary tool for all tasks.
For work with simple answer criteria—like short translations, simple summaries, sentence tone adjustments, or idea brainstorming—general models may be faster and sufficient. Using a reasoning model might only add time and cost.
The true value of this model emerges in problems with many conditions, diverging options, and high costs for mistakes.
When reasoning models fit:
- Analyzing complex code bugs
- Design decisions with tangled conditions
- Planning that must satisfy long requirements
- Reviewing the logical structure of writing
- Tasks involving math, logic, and reasoning problems
- Comparing conflicting options
A reasoning model is less like a better model and more like a tool suited for harder problems.
Understanding Tool Premises
Prompting techniques only work properly when built on an understanding of the model.
A method that was useful for general models doesn't automatically fit reasoning models the same way. Some models need more structure; others need less interference.
When using a reasoning model, what matters isn't designing the thinking process for it. It's being clear about goals, providing judgment criteria, fixing output format, and stepping back.
Trying to teach this model how to think can get in the way.
Knowing it's already built to think is what comes first.