This is a scannable reference for prompting DeepSeek-V3.2, the free, open-source (MIT) flagship released December 2025 with a 128K-token context window and quality comparable to GPT-5-class models. It runs at chat.deepseek.com. Bookmark this page and copy straight from the tables and prompt blocks below.
DeepSeek's strengths are hard reasoning and coding, and it responds best to a real problem with rich context rather than rigid formatting. Two toggles sit above the input box: DeepThink (shows the chain-of-thought for maximum accuracy) and Search (live web results). If you want fully written prompts instead of a reference, jump to our best DeepSeek prompts roundup.
1. The universal prompt formula
Every strong DeepSeek prompt follows one shape: [Role] + [Task] + [Context] + [Constraints] + [Output format]. You do not need all five every time, but the more real context you give, the better DeepSeek reasons.
| Part | What it sets | Example fragment |
|---|---|---|
| Role | The expertise DeepSeek adopts | "You are a senior backend engineer." |
| Task | The one action to perform | "Find the bug causing this timeout." |
| Context | Background, inputs, real data | "Node 20, Postgres 16, code pasted below." |
| Constraints | Limits, must-do, avoid, length | "No new dependencies, keep it under 40 lines." |
| Output format | Exact shape of the answer | "Return the fix as one code block, then a 2-line note." |
You are a senior backend engineer. Find the bug causing intermittent timeouts in the function I paste below. Context: Node 20, Postgres 16, the timeout only happens under load. Constraints: no new dependencies, keep the fix under 40 lines. Think step by step, then verify your fix against the described symptom before answering. Output the corrected code as one block, followed by a two-line note on the root cause. Code: [PASTE CODE]Best for: any analysis, coding, writing, or planning task where you can describe the problem and picture the ideal answer's shape.
2. Modes & toggles
Pick the mode and toggles to fit the task. The default chat handles most work, but DeepThink, Search, and file upload each unlock a different capability.
| Mode / toggle | What it does | When to use |
|---|---|---|
| Instant | Fast replies with no visible reasoning | Quick questions, drafting, everyday chat where speed matters |
| Expert | Deeper reasoning before it answers | Nuanced problems that need more careful thought than Instant gives |
| DeepThink | Toggle that shows the full chain-of-thought step by step | Hard reasoning, math, logic, debugging — when you want maximum accuracy and want to see the working |
| Search | Toggle that pulls live web results in real time | Current facts, recent news, anything time-sensitive; ask it to cite |
| File upload | Paperclip icon; reads PDF, DOCX, TXT, and code files | Summarizing, analyzing, or asking questions about a document or codebase |
Best for: deciding in one glance whether a prompt needs visible reasoning (DeepThink), live data (Search), or a file (paperclip).
3. Reasoning modifiers
Add one of these lines to push DeepSeek to reason carefully and check itself. They pair especially well with DeepThink turned on. More depth in how to prompt DeepSeek for reasoning.
| Phrase to add | What it does |
|---|---|
| "Think step by step." | Forces an explicit chain of reasoning before the answer |
| "Show every step of your work." | Exposes intermediate steps so you can audit the logic |
| "Define your variables first." | Pins down terms and assumptions before it calculates |
| "Consider one alternative method and compare." | Guards against a single flawed approach by cross-checking |
| "Verify against the constraints before answering." | Makes it re-read your rules and confirm the answer fits them |
| "Rate your confidence and flag any assumptions." | Surfaces shaky steps instead of a falsely confident answer |
Best for: math, logic puzzles, debugging, and any answer you plan to act on and need to trust.
4. Role modifiers
Open with a role to set the expertise and voice DeepSeek uses. Be specific — "senior Python engineer" beats "expert".
| Role phrase | Best for |
|---|---|
| "You are a senior [LANGUAGE] engineer." | Debugging, refactoring, code review, architecture calls |
| "You are a sharp copy editor." | Tightening prose, fixing tone, cutting filler without losing meaning |
| "You are a data analyst." | Reading numbers, spotting trends, building comparison tables |
| "You are a patient tutor." | Explaining a hard concept simply, step by step, with examples |
| "You are a pragmatic strategist." | Business plans, prioritization, blunt trade-off analysis |
| "You are a research analyst." | Synthesizing sources into a briefing (pair with Search) |
Best for: steering the level of detail and vocabulary so the answer matches your audience.
5. Output-format modifiers
Add one line to control how DeepSeek returns the answer. These stack cleanly on top of any prompt above.
| Phrase to add | Result |
|---|---|
| "Output in clean Markdown." | Headed sections and lists you can paste anywhere |
| "Return a table with columns X, Y, Z." | Structured comparison instead of prose |
| "Give a numbered list of steps." | An ordered checklist you can follow in sequence |
| "Output valid JSON only, no prose." | Machine-readable output for a downstream script |
| "Summarize in one paragraph." | A tight answer with no headings or bullets |
| "Walk me through it step by step." | A teaching-style, worked-through explanation |
Best for: reusing a good prompt across contexts — keep the core, swap the last line to change the shape of the output.
6. Tone modifiers
One tone line changes the whole feel of the reply. Add it near the end of the prompt so it does not get lost.
| Phrase to add | Effect |
|---|---|
| "Keep it plain and direct." | Cuts hedging and filler; gets to the point |
| "Be blunt about weaknesses." | Honest critique instead of reflexive agreement |
| "Explain it like I'm new to this." | Simple language, defined terms, worked examples |
| "Formal, no contractions." | Professional register for reports and docs |
| "Keep it warm and encouraging." | Supportive framing for coaching or feedback |
| "No hype words, no clichés." | Strips marketing fluff from the output |
Best for: matching the reply to where it's going — a Slack message, a client report, or a study session.
7. Putting it together: 5 example prompts
Five ready-to-run prompts that combine the modifiers above. Edit the bracketed parts and send. For a coding-specific set, see DeepSeek prompts for coding.
Reasoning (DeepThink on)
You are a patient tutor. Solve this problem and teach me the method: [PASTE PROBLEM]. Define your variables first, then think step by step and show every step of your work. Consider one alternative method and confirm both give the same answer. Verify the result against the constraints before answering. Output a numbered worked solution, then a one-line summary of the key idea.Note: turn on DeepThink so you can watch the reasoning and catch any slip early.
Coding
You are a senior [LANGUAGE] engineer. Here is my code: [PASTE CODE]. Task: find the bug causing [SYMPTOM]. Think step by step, then verify your fix against the symptom before answering. Constraints: keep it idiomatic, add no new dependencies. Output the corrected code as one block, then a two-line note on the root cause and how to prevent it.Note: paste the real code and the exact error — DeepSeek reasons far better with concrete inputs than with a description.
Research (Search on)
You are a research analyst. Give me a briefing on [TOPIC] as of this month. Cover the current state, the 3 biggest recent developments, key players, and open questions. Cite every claim with a link and prefer recent primary sources. Output a bulleted brief under 400 words, then a one-line bottom line. Keep it plain and direct.Note: turn on the Search toggle so DeepSeek pulls live results instead of answering from training data.
Writing
You are a sharp copy editor. Rewrite the text below to be clearer and more direct without losing meaning: [PASTE TEXT]. Keep my key terms, cut filler, and fix weak verbs. Target a confident, plain tone with no hype words or clichés. Return the rewrite first, then a short bullet list of the main changes you made.Note: asking for the change list separately keeps the rewrite clean while still showing its work.
Analysis
You are a data analyst. Compare the options in the data I paste below for a [DECISION MAKER] choosing based on [PRIORITY]: [PASTE DATA]. Think step by step and define any assumptions. Verify your ranking against the stated priority before answering. Output a table with columns Option, Strength, Weakness, Best for, then a one-line recommendation. Be blunt about weaknesses.Note: upload a spreadsheet or paste the raw rows so the comparison reflects your actual numbers.
Want the full set of written-out prompts? Bookmark the roundup at /best-deepseek-prompts and keep this cheat sheet open beside it.
Frequently Asked Questions
Which DeepSeek model do these prompts target?
DeepSeek-V3.2, the flagship released in December 2025, with a 128K-token context window. It is a free, open-source (MIT) assistant at chat.deepseek.com and is comparable to GPT-5-class models. Everything here also works on earlier DeepSeek releases.
What is the difference between DeepThink, Instant, and Expert?
Instant is the fast mode that answers without showing reasoning; Expert reasons more deeply before replying. DeepThink is the toggle above the input box that exposes the full chain-of-thought so you can see every step. Turn on DeepThink for hard reasoning, math, and anything where you want maximum accuracy.
How do I trigger DeepSeek's real-time Search?
Turn on the Search toggle above the input box, then ask for current facts and add "cite your sources with links". Search pulls live web results instead of answering from the training data, so use it for anything recent or time-sensitive.
What is the DeepSeek prompt formula?
Role + Task + Context + Constraints + Output format. Tell DeepSeek who to be, what to do, the background it needs, the limits, and the exact shape of the answer. Give it a hard problem with rich context rather than rigid formatting.
Can I over-format a DeepSeek prompt?
Yes. DeepSeek does best with a hard problem plus real context, not a wall of rigid rules. Paste the actual code, data, or text, ask it to think step by step and verify against the constraints, and keep formatting instructions to what you actually need.
Is DeepSeek free to use?
Yes. DeepSeek is a free, open-source (MIT) assistant at chat.deepseek.com. DeepThink, Search, and file upload are all available at no cost, and the model weights are openly licensed so you can also self-host.
How do the [BRACKET] placeholders work?
Anything in square brackets is a swap-in. Replace [TOPIC], [LANGUAGE], or [AUDIENCE] with your own details before sending. The rest of each prompt is written to be copy-paste ready.