How to prompt reasoning models: what changes with “thinking” AI

How to prompt reasoning (thinking) models in ChatGPT, Claude, Gemini and DeepSeek: what to keep, what to stop doing, and when a reasoning model is worth it.

By the ThinkWithPrompt editorial team at Synottic · Updated 16 October 2026

Reasoning models, sometimes called “thinking” models, work through a problem internally before they answer, so prompting them is less about telling them how to think and more about giving them the full picture: the goal, the context, the constraints and what a good answer looks like. Examples include OpenAI’s reasoning models (such as the “Thinking” modes in ChatGPT), Claude with extended thinking, Gemini’s thinking models and DeepSeek-R1. The core of good prompting still applies, but a few habits from older models are no longer needed.

What is different about a reasoning model?

A standard language model starts writing its answer straight away. A reasoning model first spends extra computation planning, trying approaches and checking its work, then writes the answer. That is why it is slower and often better on multi-step problems. It also means the techniques we used to force step-by-step thinking, such as chain-of-thought prompts, are mostly built in. What the model still cannot do is know your situation, your standards or your definition of “done” unless you tell it.

What should you keep doing?

State the goal and the decision it serves. “I need to choose between two suppliers by Friday; the main risk is delivery reliability” focuses the reasoning far better than “compare these suppliers”.
Give all the context and material. Paste the data, documents and constraints. Reasoning is only as good as what it reasons about.
Define success. What would a great answer include? What would make it wrong? Which trade-offs matter most?
Specify the output. Length, structure and audience still need to be stated; a reasoning model will otherwise decide for you.
Ask for checks you can verify. Assumptions listed, sources named, calculations shown in the final answer.

What should you stop doing?

Dropping in “think step by step”. It is redundant. If you want the steps, ask for them in the final answer (“show the calculation”).
Over-scripting the method. A rigid ten-step procedure can stop the model from finding a better route. Describe the outcome and the constraints; let it plan.
Long worked reasoning examples. Show the format you want, not the thought process.
Using a reasoning model for everything. For simple tasks it is slower and costs more without improving the result.

A prompt for a reasoning model: before and after

Habit from older models	Better for a reasoning model
You are a pricing expert. Think step by step. First list the costs, then calculate the margin, then compare to competitors, then recommend a price.	I sell handmade candles online and need to set a price for a new 300 g candle. Goal: at least 55% gross margin while staying competitive. Costs: [COSTS]. Competitor prices: [PRICES]. My customers care most about [WHAT]. Recommend a price and one alternative, show the margin calculation for each, and list the assumptions that would change your answer.

When is a reasoning model worth the wait?

Use a reasoning model	A standard model is fine
Analysing data or a complex document	Rewriting or shortening text
Planning with many constraints	Short emails and messages
Maths, logic, code and debugging	Brainstorming lists of ideas
Decisions with trade-offs	Summaries of short texts
Checking another answer for errors	Simple factual questions

Whatever the model, the answer is still a draft for you to check. Reasoning models make fewer slips on multi-step problems, but they can still be confidently wrong, especially about facts they cannot see. For a deeper look at the techniques reasoning models build in, read few-shot vs chain-of-thought prompting.

Frequently asked questions

Should I still write “think step by step” for a reasoning model?

Usually not. Reasoning models already work through a problem before answering, so the phrase adds little and can make some models produce longer, not better, answers. Spend the words on context, constraints and what a good answer looks like instead.

When should I use a reasoning model instead of a standard one?

Use one for multi-step problems where accuracy matters more than speed: analysis, planning, maths, code, tricky decisions and checking work. For quick rewrites, short emails and simple questions, a standard model is faster and usually just as good.

Do examples (few-shot prompting) still help?

Examples of the output format still help. Long worked examples of how to reason can constrain a reasoning model more than they help; try the prompt without them first, and add a single example only if the format keeps coming out wrong.

Can I see a reasoning model’s thinking?

Some tools show a summary of the reasoning, others hide it. Either way, judge the final answer on its own merits and verify important facts; visible reasoning can still contain mistakes.

Prompts that work well with reasoning models

Make AI ask me the questions that matter before it answersMakes AI identify missing information and ask the fewest questions needed before producing a high-quality result.
Answer my question, then critique it and argue against itGets an answer, an honest self-critique, the strongest counterargument, and a revised answer you can trust more.
Check AI output for errors and hallucinationsA verification checklist to catch fabricated facts, faulty logic, and unsupported claims before you rely on AI output.
Score an AI output against the prompt that produced itA requirement-by-requirement scorecard for an AI output, separating prompt problems from output problems, with prompt fixes.
Make an AI coding assistant plan a change before it edits codeA structured request that makes the assistant restate requirements, list assumptions to check, and propose file-by-file changes first.
Design an AI agent or automated AI workflowDefines an agent's goal, tools, steps, guardrails, and human checkpoints before you build it.