Prompts that play to DeepSeek's reasoning strengths

What the DeepSeek preset adds

For DeepSeek, the generator uses Markdown sections and adds reasoning guidance that other presets do not get in Advanced mode: a line asking the model to think the problem through step by step before writing, then present only the polished final answer. In Expert mode it also asks DeepSeek to reason carefully and show the final answer clearly separated, which makes long outputs easier to scan and copy. Qwen, Llama and Mistral get the same step-by-step line because they benefit from it in similar ways.

An Expert-level prompt for a coding task looks like this (shortened):

## Role
You are a senior software engineer who writes clean, secure, well-documented code.

## Context
Flask app, Python 3, SQLite. Users report duplicate orders when they double-click "Pay".

## Task
Debug the duplicate-order bug in my Flask checkout and propose a fix.

## Requirements
- Explain the approach before the code
- Handle edge cases and errors

## Output format
Complete, runnable code in one block, followed by a short explanation.

## Quality check
- Every requirement above is met
Think through the problem step by step before writing, then present only the polished final answer.
Reason carefully step by step; show the final answer clearly separated.

Chat mode or reasoning mode?

The DeepSeek app offers a regular chat mode and a reasoning mode (labelled DeepThink at the time of writing), plus a web search option. The reasoning mode works through the problem at length before answering, which is slower but stronger on maths, logic, debugging and multi-step planning. Regular chat is faster and better suited to writing, rewriting and quick questions.

With reasoning mode, the model already thinks step by step, so the reasoning line in the prompt is redundant but harmless; you can delete it to keep the prompt short. What matters more for reasoning models is a precise goal, the constraints that define a correct answer and the format of the result. Detailed instructions about how to think usually help less than a clear description of what a good answer looks like.

Goals and constraints beat long instructions

Guides on prompting reasoning models, including DeepSeek's own notes on its reasoning model, have suggested putting all instructions in the user message and preferring zero-shot prompts (no examples) because examples can narrow how the model approaches the problem. Treat that as a starting point and test both ways for your task. In practice:

• Give the inputs in full: the error message, the table, the exact numbers.
• State success criteria: "the fix must not change the database schema", "the answer must be an integer".
• Ask for the format of the final answer only, not the format of the thinking.
• For maths, ask it to put the final result on its own line so you can find it quickly.

Check the answer, especially for code and numbers

Strong reasoning still produces mistakes, and confident step-by-step explanations can make them look convincing. Run the code, recheck key calculations and ask DeepSeek to list its assumptions; the Expert quality check asks it to flag uncertain facts. For long code or documents, estimate the size first with the token counter so you know whether everything fits in one message. To compare formats across assistants, see the Claude prompt generator or start from the universal AI prompt generator.