Chain of thought and reasoning models

The concept

Chain-of-thought prompting used to mean asking the model to show a step-by-step reasoning trace. Current reasoning models make that pattern less reliable as a generic improvement technique.

There is a useful distinction: ask for the answer, assumptions, evidence, and tradeoffs you need to review. Do not ask for private scratch work just because an old template did.

The current approach

Use provider controls for reasoning depth when they exist. OpenAI documents reasoning models and separate reasoning best practices. Claude and Gemini expose their own thinking features and controls.

For a practical decision prompt that uses this idea, see decision-making and tradeoff analysis.

In the prompt, ask for reviewable output:

Recommend one option. Include the assumptions that would change the recommendation,
the evidence you relied on, and the main risk in the chosen path.

That is different from “think step by step.” It asks for the parts a human can use.

When visible reasoning helps

Visible reasoning is useful when the reasoning itself is the deliverable. Tutoring, math explanations, code walkthroughs, and decision memos may need a clear explanation.

Even then, ask for the explanation shape you need. “Show the derivation a beginner should see” is better than “reveal all chain of thought.”

What to avoid

• Adding “think step by step” to every prompt.
• Asking for hidden reasoning when a short justification would do.
• Confusing a longer explanation with a better answer.
• Telling a reasoning model exactly how many internal steps to take.