How to Prompt Reasoning Models (Without Getting in Their Way)

A strange thing happened as AI models got better at thinking: a lot of the prompting advice that made them better started making them worse. Think step by step, carefully staged reasoning scaffolds, piles of examples — techniques that were genuinely helpful on earlier models can degrade the output of modern reasoning models. If your hard-won prompting habits suddenly feel like they're backfiring, this is probably why.

Here's what changed, and how to prompt this new class of model well.

What a reasoning model actually is

Reasoning models — the thinking variants like Claude's extended thinking, OpenAI's o-series, Gemini's thinking mode — do something ordinary chat models don't: they reason internally before they answer. Given a problem, they work through it step by step on their own, privately, and then give you the result of that thinking.

That single fact changes everything about how you should prompt them, because the thing you used to do manually — coax the model into reasoning out loud — now happens automatically. When you try to do it anyway, you're not helping. You're interrupting.

The mental model that fixes most mistakes

Chat models reward you for thinking out loud with them. Reasoning models reward you for stating the problem clearly and getting out of the way.

Internalize that and most of the specific do's and don'ts follow naturally. The old style was collaborative narration — walk the model through the reasoning. The new style is clean delegation — describe the destination and let the model find the route.

What to stop doing

Stop telling it to think step by step. The model already is. At best the instruction is redundant; at worst it disrupts the reasoning process the model would have run on its own. This is the single most common mistake, precisely because it used to be the single best trick.

Stop over-stuffing context and examples. This is counterintuitive, so it bears emphasizing: reasoning models can overthink. Hand them a mountain of few-shot examples and verbose background and they take longer and produce worse answers — the extra material redirects or overwhelms the internal reasoning instead of guiding it. On reasoning-heavy tasks, examples can actively bias the model toward copying surface patterns rather than reasoning from scratch. (See Few-Shot vs Zero-Shot for exactly when examples help versus hurt.)

Stop building elaborate reasoning scaffolds. Multi-stage first do this, then consider that, then weigh the following structures were a way to externalize thinking the model couldn't do alone. A reasoning model can do that structuring itself, usually better than your scaffold.

What to do instead

State the problem and the goal, then stop. A clean description of the situation plus a clear picture of what a good outcome looks like is the ideal reasoning-model prompt. Give it the what and the why; let it own the how.

Trim your context to what's genuinely relevant. Since extra material can degrade reasoning, curation matters more, not less. Include the facts the problem actually needs and cut the rest. Quality of context beats quantity.

Be clear about the output you want. Get out of the way applies to the reasoning, not the requirements. You should still be specific about format, length, and constraints on the final answer — that's not interfering with thinking, it's defining the target.

Let it work. Reasoning models often take longer to respond because they're actually reasoning. That latency is the feature. If you're using one through an API, this is also exactly the situation where streaming and generous timeouts matter.

When you don't want a reasoning model

Worth saying: reasoning models aren't the right choice for everything. For simple, fast, high-volume tasks — classification, extraction, straightforward rewriting — a standard model is cheaper, faster, and every bit as good. The extra thinking is wasted effort on a task that doesn't require thought. Match the model to the job: reserve reasoning models for problems that genuinely have reasoning in them.

The takeaway

Prompting a reasoning model is an exercise in restraint. The instinct built up over years of using chat models — add more guidance, spell out the steps, show more examples — is exactly backwards here. State the problem cleanly, give it the relevant context and nothing more, be clear about the output, and trust the model to think. The best thing you can do for a model that reasons well is stop trying to reason for it.

This is one of the biggest shifts covered in The Complete Guide to Prompt Engineering in 2026. For the flip side — the tasks where showing examples still pays off — see Few-Shot vs Zero-Shot. And browse the prompt library for prompts you can adapt.