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 messageR1 ignores the system role; instructions must go in user
Do not use few-shot examples; let the model reason aloneR1 reasons better unaided; few-shot distracts
Set temperature to 0.6R1 official recommendation, balances stability and diversity
Give the question directly; trust the chain-of-thoughtNo "think step by step" needed; R1 thinks internally
Take the final answer, ignore the reasoning traceR1 emits a long trace; use the last conclusion
V3 Chat Rules (deepseek-chat) 5
Use a system prompt to lock role and behaviorV3 supports system for tone-setting
Few-shot works — give samples for complex formatsV3 mimics the example structure
Creative tasks temperature 0.7-1.0, precise 0.3-0.5Tune by task type
top_p 0.95V3 default nucleus value
Enable JSON mode / function calling for structured outputV3 supports native structuring
Code Generation 5
V3: implement {req} in {lang} with error handling, comments, unit testsV3 is good for direct codegen
R1: reason about the optimal {algorithm} and explain complexityR1 suits algorithm/hard-logic inference
Specify framework and version to avoid hallucinated APIsExplicit constraints cut hallucinations
Ask for a runnable snippet + usage exampleEasy to verify directly
Review generated code for safety before useAI code needs human sign-off
Chinese & Long Text 5
Prompt in Chinese directly; no need to translate to EnglishDeepSeek is strong in Chinese
128K context: chunk long docs and cite sourcesChunking is more stable for very long text
Summarize/extract/translate with V3V3 has good throughput and instruction-following
Use R1 for deep analysis with enough contextR1 suits in-depth reasoning
Ask the model to cite paragraph locationsEasier to verify long-text claims
Templates 5
You are {role}, goal {goal}, constraints {constraints}. Task: {input}V3 role+goal+constraint triplet
Reason: {question}. Give a conclusion and show the derivation.R1 reasoning-chain template
Compare A/B in a Markdown table, dimensions: {dimensions}V3 comparison template
Output strictly per this JSON Schema, no explanation: {schema}V3 structured JSON template
From the material above, list 3 pros/cons with evidenceV3 multi-angle analysis template
Pitfalls 5
R1: no few-shot / no systemWeakens reasoning quality
Precise tasks: avoid high temperature0.3-0.5 is steadier
Watch truncation on long outputs; split if neededAsk in batches for very long results
R1 reasoning trace is not the final answerUse only the conclusion section
For sensitive/factual asks, have the model flag uncertaintyReduce hallucination risk
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.
Official References
Commands are compiled from the official docs below. Click to verify the latest usage.
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