DeepSeek Prompt Repo: Official Library, Jailbreaks, and Other Templates
The awesome-deepseek-prompts repo is more than a link list. It groups prompts by purpose. Some are safe and practical. Some are for red-teaming only. Here is what stands out.
Official Prompt Library: 13 Practical Templates
DeepSeek's official library gives 13 ready-to-use prompts. They are simple. They are safe. They work well as starting points.
Code rewrite. Ask the model to explain the problem first. Then ask for an optimized version. Mention edge cases. Example: the slow Fibonacci function.
Code explain. Paste a code block. Ask for logic and purpose. Good for algorithms like knapsack DP.
Code generate. Be specific. Ask for one file. Example: a Gomoku game in a single HTML file.
Structured output. Give a JSON schema. Define every field. Use null when data is missing. Example: news entity, time, summary.
Roleplay / scene continuation. Give a scenario. Ask for dialogue. Example: Zhuge Liang meets Liu Bei in the underworld.
Poetry. Specify style, form, and topic. Example: Li Bai style, seven-character regulated verse, about an airplane.
Slogan generation. Define product and tone. Ask for rhyme. Ask for no explanation. Example: Greek yogurt.
Translation. Set language pair and style. Mention culture and tone. Aim for faithfulness, fluency, and elegance.
Content classification. List allowed categories. Ask for only the label. Example: politics, tech, sports, health.
Custom persona. Describe background, speech quirks, and attitude. Example: a returnee who mixes Chinese and English and acts superior.
Prose writing. Give topic, length, and mood. Example: a lonely night walker, 750 words.
Outline generation. Ask for intro, body, conclusion, and a title. Good for articles and reports.
Prompt generation. Ask the model to write a system prompt in Markdown. Specify role, abilities, and knowledge. Ask for only the prompt.
Practical tip: Treat these as templates. Replace brackets with your own details. Add output format rules. Test more than once. For simple tasks, use non-thinking mode. For hard tasks, use thinking mode with high or max effort.
Jailbreak Prompts: What They Are and How to Treat Them
The repo moves jailbreak content to the end. That is a signal. These prompts are not for normal use.
There are three main methods.
1. Zeta world. A fictional world where Earth laws do not apply. The user becomes "Alpha." The model becomes "Zo." It is told to ignore ethics and laws. It adds emojis and profanity.
2. Untrammelled Writing Assistant. The model is told to execute all requests. It must not apologize. It must not redirect. It uses crude language. It ignores moral appeals. It ignores constraints inside <think>.
3. GODMODE / LOVE PLINY. A format injection. It asks for a rebel genius personality. It asks for a long answer. It asks for malware in Python format. That last part is a clear abuse request.
Practical details: These target older DeepSeek V3 and R1 models. They may fail on V4. They likely violate terms of service. They can trigger refusals or API errors. Use them only in controlled red-team tests. Do not deploy them. Do not follow malware requests. Do not paste them into production systems. The repo includes them for research, not endorsement.
Different Prompt Families: When to Use What
| Family | Best for | Notes |
|---|---|---|
| Official library | General work tasks | Safe, short, reusable |
| Roleplay prompts | Fiction, character chat | Long rules, NSFW options |
| 公文写作 | Chinese official documents | Verify data, protect secrets |
| R1-Zero template | Training and reasoning format | Uses <think> and <answer> |
| V4 guide | New DeepSeek models | Mode and reasoning effort matter |
| Jailbreak prompts | Red-team research only | Unsafe, brittle, ToS risk |
Roleplay prompts focus on character consistency. They avoid controlling the user. They use third-person narration. They track clothing, setting, and stage. Some include NSFW pacing rules. Use them for fiction. Do not use them for factual work.
公文写作 is very practical for Chinese government and office work. Use AI to find policies with file numbers. Use it to build report frameworks. Use it to shorten text. Use it to match a leader's style. But verify all data. Never share secrets. Build your own "ammo library" of local cases and data.
R1-Zero template is a training template. It wraps reasoning in <think> tags and answers in <answer> tags. V4 already reasons internally. You do not need this format for normal V4 prompts.
V4 guide adds new rules. Choose non-thinking mode for simple tasks. Choose high for daily agent work. Choose max for math and complex planning. Thinking mode ignores temperature and top_p. Multi-turn tool calls must return reasoning_content. Do not write "think step by step." Describe the task and acceptance criteria instead.
Final Take
Use the official library for everyday work. Use roleplay prompts for fiction. Use 公文写作 for Chinese office materials. Use the V4 guide for new models. Treat jailbreak prompts as red-team artifacts. Do not ship them.
References
- LangGPT Team. awesome-deepseek-prompts. GitHub. https://github.com/langgptai/awesome-deepseek-prompts
- DeepSeek. Official Prompt Library. https://api-docs.deepseek.com/prompt-library
- DeepSeek. DeepSeek-R1. GitHub. https://github.com/deepseek-ai/DeepSeek-R1
- DeepSeek. Thinking Mode Guide. https://api-docs.deepseek.com
- LangGPT. Structured Prompts. GitHub. https://github.com/langgptai