DeepSeek Prompts Cheatsheet - R1 Reasoning & V3 Chat Writing Techniques
Essential DeepSeek prompt templates organized by model — R1 reasoning rules, V3 chat tricks, codegen, structured output. Copy a template, fill the variables, and ship.
R1 Reasoning Rules (deepseek-reasoner) 5
No system prompt — put all instructions in the user messageDo not use few-shot examples; let the model reason aloneSet temperature to 0.6Give the question directly; trust the chain-of-thoughtTake the final answer, ignore the reasoning traceV3 Chat Rules (deepseek-chat) 5
Use a system prompt to lock role and behaviorFew-shot works — give samples for complex formatsCreative tasks temperature 0.7-1.0, precise 0.3-0.5top_p 0.95Enable JSON mode / function calling for structured outputCode Generation 5
V3: implement {req} in {lang} with error handling, comments, unit testsR1: reason about the optimal {algorithm} and explain complexitySpecify framework and version to avoid hallucinated APIsAsk for a runnable snippet + usage exampleReview generated code for safety before useChinese & Long Text 5
Prompt in Chinese directly; no need to translate to English128K context: chunk long docs and cite sourcesSummarize/extract/translate with V3Use R1 for deep analysis with enough contextAsk the model to cite paragraph locationsTemplates 5
You are {role}, goal {goal}, constraints {constraints}. Task: {input}Reason: {question}. Give a conclusion and show the derivation.Compare A/B in a Markdown table, dimensions: {dimensions}Output strictly per this JSON Schema, no explanation: {schema}From the material above, list 3 pros/cons with evidencePitfalls 5
R1: no few-shot / no systemPrecise tasks: avoid high temperatureWatch truncation on long outputs; split if neededR1 reasoning trace is not the final answerFor sensitive/factual asks, have the model flag uncertaintyCommand Examples
R1 reasoning template: ask directly, no system message
Reason through the following and state both your conclusion and the derivation behind it: {question}Every instruction sits in the user message, which is what R1 requires given it ignores the system role; the reply opens with a long derivation, so read the concluding section.
V3 three-part template: role, goal, constraints
You are {role}. Your goal is {goal}. Your constraints are {constraints}. Task: {input}The three parts work because each occupies a fixed slot; on V3 you can also park this line in a system prompt for reuse and keep only the task in the user message.
V3: strict JSON Schema output, no commentary
Emit output that strictly conforms to the following JSON Schema, with no explanation: {schema}Structured data belongs on V3 with its JSON mode or function calling, which is steadier than asking R1 for JSON; the no-explanation clause keeps stray prose out of the payload.
V3: comparison table plus multi-angle analysis
Compare A and B in a Markdown table across these dimensions: {dimensions}
Based on the material above, list three arguments for and against, each with its supporting evidence.Two instructions in one request: the table compresses the conclusion into something scannable, then the second half forces the other side of the argument so the answer is not one-sided.
Common Pitfalls
- Writing a system prompt or few-shot examples for R1: the former is ignored and the latter biases the chain, so reasoning quality drops instead of rising.
- Shipping the middle of an R1 reply as the answer: that section is its own scratch work, and the conclusion sits at the end.
- Leaving temperature above 0.7 for exacting work such as extraction or conversion; the 0.3-0.5 range is the stable one.
- Expecting an unbounded long reply in a single pass: once it truncates, the reasoning you paid for is gone, so ask for the remainder in batches.
- Assigning structured JSON extraction to R1: it is not tuned for it, the schema drifts, and the job belongs on V3 with JSON mode.
Tips
- The key DeepSeek difference: R1 (reasoning) and V3 (chat) have opposite prompt rules — R1 forbids system/few-shot, V3 uses system to lock behavior and few-shot works. Mismatching model and rule drops quality sharply.
- Prefer V3 for codegen (runnable snippets) and R1 for algorithm/hard-logic inference; pick the model by task.
- For Chinese prompts just use Chinese — DeepSeek is as strong in Chinese as in English, no translation needed.
- For structured data use V3 JSON mode + function calling; more stable and controllable than asking R1 for JSON.
FAQ
What exactly differs between prompting R1 and prompting V3?
It comes down to where instructions live and whether you supply examples. R1 (deepseek-reasoner) takes no system prompt and should get no few-shot examples: everything goes in the user message. V3 (deepseek-chat) is the mirror image: a system prompt is the right tool for pinning role and behavior, and few-shot examples for tricky formats get copied. Handing V3-style prompts to R1 makes the instruction fail silently.
Do I need to tell R1 to think step by step?
No. R1 runs its own chain of thought, so the phrase only adds noise; the guidance is to hand it the question and trust the internal reasoning. The corollary matters just as much: the long reasoning block in the reply is not the answer, and what you want is the concluding section at the end.
How should I set temperature and top_p?
R1 has a recommended value of 0.6 that already balances stability against variety, so leave it there. For V3, pick by task: 0.7-1.0 for creative work, 0.3-0.5 for precise work, with top_p at its 0.95 default. And on the R1 side the real mistake is not the temperature at all: cramming in few-shot examples and a system prompt degrades the reasoning rather than sharpening it.
R1 or V3 for code?
The split is clean. V3 produces runnable code directly, which covers ordinary implementation work with error handling, comments, and unit tests. R1 is for algorithms and hard reasoning, where you ask for an optimal implementation plus its complexity. Three habits apply either way: pin the framework and version so it cannot invent APIs, ask for a runnable snippet with a usage example so you can verify it, and review the result for security before it lands.
Anything special about long documents and Chinese-language prompting?
Ask in Chinese directly, no need to translate. Once you are working near the 128K window, still chunk very long material and label the sources, and ask for the paragraph position behind each conclusion so you can check it. The division of labor: summarizing, extraction, and translation go to V3 for throughput; deep analysis of a whole long document goes to R1, and there you supply the full context before posing the question.
Official References
Each command links to its official documentation below, so you can verify the latest usage and read deeper.
Maintained by LaoHand
Publicly updated on Aug 21, 2026, continuously proofread against official docs.
Contact Us
Wrong command or description? Send us corrections, business inquiries or product feedback by email.
Contact Us