Prompt Patterns
Reusable DeepSeek prompt templates for summarization, coding, FAQs, SEO content, and knowledge tasks.
DeepSeek Prompt Patterns
The goal of prompting is not to be clever.
The goal is to be repeatable, reviewable, and useful.
Pattern 1: long-document summarization
Read the following material and return a structured summary.
Audience:
Product manager / decision maker
Goal:
Extract only the information that changes decisions.
Output requirements:
1. 5 key takeaways
2. Key facts or data points
3. Risks
4. Unanswered questions
5. Flag obvious contradictions if present
Material:
{{context}}Pattern 2: code review
Review the following code as a senior engineer.
Goal:
Find issues that affect reliability, performance, maintainability, and system boundaries.
Output format:
1. Overview
2. High-priority issues
3. Suggested fixes
4. Side effects or tradeoffs
5. Regression test ideas
Code:
{{code}}Pattern 3: FAQ generation
Build an FAQ from the following material.
Requirements:
1. Return 10 Q&A pairs
2. Questions must be written from the user's perspective
3. Answers should be concise but complete
4. Mark unknown points as "needs human confirmation"
Material:
{{docs}}Pattern 4: SEO rewriting
Rewrite the following content around the keyword {{keyword}}.
Target page:
Homepage hero / category intro / FAQ / blog summary
Requirements:
1. Keep the core information
2. Remove filler
3. Write naturally without keyword stuffing
4. Add clear subheadings
5. Return a 120-160 character search snippet
Source content:
{{content}}Pattern 5: knowledge-grounded Q&A
Answer only from the material I provide.
Requirements:
1. Give the conclusion first
2. Then show the supporting basis
3. If the source is insufficient, say "insufficient source material"
4. Do not invent unsupported facts
Question:
{{question}}
Material:
{{context}}The four most common prompting mistakes
1. The goal is vague
"Take a look at this document" is worse than "return five decision-relevant conclusions for a PM."
2. The context is incomplete
The user knows the background. The model does not.
3. The output format is undefined
If you do not specify the structure, you usually get text that is harder to reuse.
4. There is no second-pass review
Many answers improve dramatically after one follow-up round asking for assumptions, missing information, or edge cases.
Practical recommendation
- Build fixed prompt templates for recurring tasks
- Test them with real business examples
- Define mandatory output fields
- Ask the model to mark uncertainty explicitly