Prompting reasoning models in 2026

Why it matters

Many prompting habits were invented for models that did not reason before answering. Current reasoning models from Anthropic, OpenAI and Google think before they respond, and several of the old tricks now do nothing, waste effort or get in the way. This criterion, model fit, asks whether a prompt suits the model it is written for. Here are the five changes that matter most in 2026.

Set the effort; do not script the thinking. Claude models use adaptive thinking, and the effort setting in the API (from low to max) is the main control for how deeply they think. OpenAI offers a reasoning-effort setting too, and its guidance for reasoning models says that asking them to think step by step is unnecessary and can sometimes hinder them. Google notes that Gemini 2.5 and 3 models think internally, so the reply does not need to spell out the reasoning. Anthropic adds that a general instruction such as “think thoroughly” often beats a hand-written step-by-step plan.

Write calmly. Capitals, exclamation marks and absolute words were once a way to make a model pay attention. Current models follow instructions closely, and Anthropic's guidance warns that aggressive wording can make them over-apply a rule; it suggests replacing a capitalised command with a plain “use this tool when…”. OpenAI's GPT-5 guide makes a related point about contradictions: the model spends effort reconciling them.

Long material first, question last. Anthropic advises placing long documents above the instructions and the question, and Google gives the same advice for Gemini 3: supply the context first and put the instruction at the end.

Explain the why. A rule with its reason (“keep it short because it goes into a client email”) lets the model generalise; a bare rule invites literal, brittle obedience. Anthropic's guidance recommends explaining the motivation behind instructions.

Use structured outputs, not prefill. Starting the assistant's reply for it (prefill) used to force JSON. On current Claude models, a prefilled final assistant turn is no longer supported, and Anthropic points to structured outputs instead. Anthropic and OpenAI both let you pass a JSON schema with the request so the response follows it.

What good looks like

• 0 — Contradictions, shouting or harmful tricks. Capitalised commands, forced step-by-step reasoning, a reply that relies on prefill, or “be brief but cover everything”.
• 1 — Calmer, but obsolete habits remain. “Think step by step. What are the risks?” followed by 20 pages of contract.
• 2 — Plain and calm, weak structure. Material and questions mixed in one paragraph, with no delimiters.
• 3 — Calm, consistent, well placed. Long material first in tags, the request last, and the reason given where it matters.
• 4 — All of that, written for the chosen model. Structure in the vendor's recommended form, depth left to the effort setting, a check the model can run on itself, and nothing redundant.

Common mistakes

• Keeping “think step by step” or “take a deep breath” out of habit.
• Using capitals or repetition to signal importance instead of a reason.
• Pasting a long document after the question.
• Relying on prefill or on “reply only in JSON” when the API offers structured outputs.
• Asking the model in the prompt to think less or more instead of changing the effort setting.

Try it on your own prompt

Paste an older prompt into the workspace, one you have reused for a while, and look for the habits above: capitals, step-by-step scripts, a question before the material, rules without reasons. Fix them one at a time and watch the model fit score. Then try optimise for your target model family and compare its version with yours.