How to prompt reasoning models (GPT-6, Claude Opus 5.5, Gemini, Grok, DeepSeek)

Practical patterns for prompting models that think before answering: what to specify, what to leave to the model, and how each current model sets reasoning depth with effort or thinking level instead of prompt wording.

← Back to guides

What makes reasoning models different

Reasoning models (GPT-6, Claude Opus 5.5, Gemini 3.x with thinking, Grok 4.7, DeepSeek in thinking mode) reason internally before producing output — they spend tokens exploring before they answer.
They handle ambiguous, multi-step tasks better than standard models, but require a different prompting style: less scaffolding from you, more clarity about what success looks like.
Explicit step-by-step prompting (like 'think step by step') is often redundant — the model already does this internally. Overloading it with process instructions can interfere with its own reasoning.

What to specify and what to leave to the model

Specify: the goal, the constraints, the output format, and what counts as a correct answer. Give it enough grounding to evaluate its own work.
Leave to the model: the reasoning path, intermediate steps, and how to decompose the problem. It handles this better than you can prescribe.
For factual or research tasks: require citations or grounded claims. Reasoning models can still hallucinate — they reason better, but they are not immune to confabulation.
For mathematical or logical tasks: ask for a final answer with a short justification, not a full narration of every step. Narrated steps can introduce errors in long chains.

How current models set reasoning depth

Depth is a request setting, not a phrase. On current models you raise or lower reasoning with a parameter (effort, reasoning_effort or thinking_level). Writing 'think harder' in the prompt is not the control.
Claude Opus 5.5 always thinks; effort sets how deeply (default medium). Drop 'think carefully' or 'double-check your work' lines and state the goal, context, constraints and what 'done' looks like in one complete request.
GPT-6 Astra needs less step-by-step instruction than earlier GPTs: give it the goal, constraints and permissions. It has no 'none' effort, so 'low' is the floor. GPT-6 Sol starts at medium effort and does better with tighter prompts: explicit steps and an exact format.
Gemini 3.8 Flash uses thinking_level (low, medium or high; default medium) and answers tersely by default, so ask for detail when you need it. Gemini 3.1 Pro defaults to high.
Grok 4.7 reasoning can't be turned off: reasoning_effort runs from low to xhigh (default high). Lower it for latency-sensitive work instead of asking for a 'quick answer' in prose.
DeepSeek Flash thinks by default (effort high). Drop to low or turn thinking off for quick tasks, and in tool loops pass reasoning_content back in full.
These defaults come from each provider's documentation as of Promptea's last registry check (2026-09-24). Parameters change, so confirm them in the provider's docs before you tune a production integration.

Common mistakes

Over-scaffolding: adding 'First, list all options. Then, evaluate each. Finally, pick the best.' — the model already does this; your instructions can conflict with its own plan.
Under-specifying the goal: vague success criteria mean the model may reason well and still give you an unusable answer.
Ignoring output format: reasoning models can produce long, unstructured answers. Always specify the format you need, or you will spend time parsing the result.

Templates

Complex multi-step reasoning

Goal: [what a correct answer looks like — one sentence]
Context: [key facts, constraints, definitions]
Output format: [bulleted recommendation / decision table / ranked list]
Quality bar:
- Must include: [specific items]
- Must avoid: [forbidden claims or assumptions]
If the answer requires trade-offs, state them explicitly.

Opens the analyzer with the prompt prefilled.

Open in Promptea

Structured analysis with justification

Analyze the following and return:
1. Final answer (1–2 sentences)
2. Key reasoning (max 3 bullets — cite the data)
3. Confidence: High / Medium / Low and why

Data:
"""[paste]"""

Constraints: [e.g. use only the provided data, no external assumptions]

Opens the analyzer with the prompt prefilled.

Open in Promptea

FAQ

Should I still write 'think step by step' for reasoning models?

No. Reasoning models run an internal chain-of-thought automatically. Adding it is at best redundant and at worst disruptive. Focus on specifying the goal and output format clearly instead.

Are reasoning models more accurate for math and logic?

Generally yes for multi-step problems, but they still make arithmetic errors and can confabulate facts. Always verify numerical results and require citations for factual claims.

How do I make a reasoning model think more or less?

Use the model's reasoning setting, not prompt wording: effort on Claude Opus 5.5 and GPT-6, thinking_level on Gemini 3.x, reasoning_effort on Grok 4.7, and effort or thinking mode on DeepSeek. In chat apps this is usually the model or mode picker. Keep the prompt focused on the goal, constraints and output format.