How do you prompt a reasoning model?

Reasoning models are now everywhere you look. OpenAI's o1 is in ChatGPT, o3-mini reached developers on 31 January, and DeepSeek's open R1 model is free to try. All of them think through a problem privately before they reply. That changes how you should ask. When OpenAI launched o3-mini for developers, it said plainly that prompting its reasoning models differs from prompting its GPT-series models, and pointed people to a separate guide.

The short version: chat models reward a conversation, while reasoning models reward a good brief. Below is what that means in practice.

From conversation to brief

With a chat model such as GPT-4o, most of us start with a short question and refine it over several turns. The model asks for what it is missing, or its weak answer shows us what to add.

Ben Hylak, founder of Dawn Analytics, wrote a much-discussed guest post for the Latent Space newsletter in January. He describes going from calling o1 garbage to using it every day once he stopped treating it like a chat partner. His conclusion is that o1 behaves more like a report generator: it takes a lazy question at face value instead of drawing context out of you, so you have to push the context in. His advice is to give it far more than feels natural, and to treat it like a new hire on their first day.

What belongs in that context:

• What you are trying to achieve and why.
• What you have already tried and what happened.
• Your constraints: budget, tools, deadlines, style rules.
• Any terms specific to your company or field, with definitions.
• The raw material itself: the code, the data, the document.

Say what, not how

Older prompting advice often told the model how to think: "You are an expert. Think slowly. First list the assumptions, then…" Hylak found the opposite worked better with o1: describe the output you want in detail, and let the model plan its own route. Reasoning is the part these models are built to do.

That makes the end of your prompt the most important part. Spell out the format, the length, what to include and what to leave out. Hylak also suggests asking for one specific output per prompt, and Latent Space's editor adds a useful tip: describe what makes a result good or bad, so the model can check its own work against those criteria.

Here is a brief built that way:

I run a two-person bookkeeping firm with about 60 small-business clients. We send monthly reminders by email, and roughly a third of clients upload their receipts late, which pushes our closing past the 10th of the month.

What we've tried: a second reminder on the 5th (small improvement), and a late fee (two clients left, so we dropped it).

Constraints: no new paid software over $30 a month; clients are mostly over 50 and use email, not apps.

Goal: a plan to get at least 80% of receipts in by the 3rd. Give me three options ranked by effort, each with the first step we'd take next Monday and one risk. A good answer is specific to a tiny firm; a bad one assumes we have staff or an IT department.

Notice there is no role and no "think step by step." There is a lot of situation, a precise goal, and a definition of success.

Model-specific settings

Each family has its own quirks:

For math with DeepSeek-R1, its makers suggest asking for the final answer inside \boxed{} so it is easy to find in a long response.

When should you still use a regular chat model?

Often. Reasoning models are slower, sometimes by minutes, and they can spend a long time on a question that needs a sentence. Hylak notes that o1 can wander into reasoning rabbit holes on very simple tasks. For quick facts, casual back-and-forth, brainstorming or a fast rewrite of a paragraph, a chat model such as GPT-4o answers sooner and is perfectly good.

Writing is a second case. Hylak found o1 poor at writing in a specific voice or style, with a pull toward an academic, report-like tone. If you want a warm newsletter intro or a message in your own voice, a chat model usually gets there with less effort, and you can steer it turn by turn.

A practical pattern is to combine the two. Use a reasoning model for the hard, high-stakes thinking: the plan, the diagnosis, the code that must work the first time. Then hand its output to a chat model to polish the wording, shorten it or change the tone. Pick the tool by the job, not by which model is newest.