These 13 DeepSeek prompts are written for what DeepSeek's own documentation says about its models: a thinking mode that reasons before it answers, a 1M-token context window, a JSON output mode with specific rules, and recommended temperatures for each kind of task. Each one is a fill-in template you can paste into chat.deepseek.com or send through the API.
General jobs like editing a draft or planning a project are in the best ChatGPT prompts and the Claude prompts, and everyday code work is in the coding prompts. Nothing here repeats them.
How to use these prompts
Copy a prompt, replace each blank in double braces, or keep the default after the colon. {{language:Python}} means Python unless you change it.
Know which model and mode you are using, because DeepSeek has changed both a lot in 2026. The V4 preview announcement of 24 April 2026 introduced V4-Pro and V4-Flash with open weights, said "1M context is now the default across all official DeepSeek services," and pointed web users to Expert Mode and Instant Mode. The V4-Pro release notes of 13 August say V4 Pro runs in Expert Mode on the app and web. The most recent changelog entry, from 17 October, upgraded V4.1 Flash and says the web and app models were updated too. The old deepseek-chat and deepseek-reasoner names were retired on 24 July 2026, so older prompt guides that tell you to pick "R1" are out of date.
Three things from the docs shape the prompts below:
- Thinking mode reasons first. The thinking mode guide says the model produces chain-of-thought before its final answer to improve accuracy. In the API it is on by default, with effort levels that the release notes describe as "low for simple tasks, high for daily Agent workflows, max for complex tasks."
- Temperature depends on the task. The parameter settings page recommends 0.0 for coding and math, 1.0 for data cleaning and analysis, 1.3 for conversation and translation, and 1.5 for creative writing. Thinking mode does not support
temperature, so these only apply with thinking off. - JSON has rules. The JSON output guide says to "Include the word "json" in the system or user prompt" and to "provide an example of the desired JSON format."
Which DeepSeek prompts suit thinking mode?
Use these with thinking on, which is the API default. They give the model a clear problem and a clear finish line, and leave the reasoning to the mode.
1. A math or logic problem with a checkable answer
For quantitative problems where you want the answer, not an essay. The fixed final line makes the answer easy to find and compare.
Solve the problem below.
<problem>
{{problem}}
</problem>
State any assumption you need. Give the final answer on its own line
in the form "Final answer: ...", with units if there are any. Then
check the answer by substituting it back or by a second method, and
say whether the check passed.
2. Design an algorithm with a correctness argument
For coding problems where the hard part is the approach, not the syntax.
Design an algorithm for this problem:
{{problem_statement}}
Constraints: {{constraints:n up to 10^6, memory limit 256 MB}}
Give, in order:
1. The idea in three sentences.
2. Why it is correct, as a short argument.
3. Time and memory complexity.
4. Edge cases and how the algorithm handles each.
5. An implementation in {{language:Python}} with a few test cases.
3. Find the flaw in an argument
Paste a proof, a business case or a design rationale. A reasoning model is well suited to checking a chain step by step.
Below is an argument I want to rely on.
<argument>
{{argument}}
</argument>
Check it step by step. For each step, say whether it follows from the
previous ones, and if not, what is missing. Identify the weakest step
and whether the conclusion still holds without it. If the argument is
sound, say so plainly instead of inventing problems.
4. Turn a slow reasoning answer into a fast prompt
Thinking mode costs time and tokens. Once it has solved a task well, use this follow-up to get a prompt that non-thinking mode can repeat on similar inputs.
You just solved this task well. I need to run the same kind of task
many times in non-thinking mode, which won't reason at length.
Write a reusable prompt that captures the method you used: the steps
in order, the checks that mattered, and the output format. Mark the
part that changes with the word INPUT in capitals. Keep it under
{{length:250 words}} and include one short worked example.
How do you use DeepSeek's 1M-token context?
With 1M tokens as the default, you can paste a whole service, a month of logs or a stack of contracts. The usual rule for long prompts still applies: material first, question last, and labels so the model can point back to where it found things.
5. Ask a question about a whole codebase
Concatenate the files with a path header before each one. The headers let the answer cite real file names.
Below are the source files of {{project:our billing service}}, each
starting with a line "=== path/to/file ===".
{{files}}
Based on the files above, answer: {{question:where is a failed payment retried, and what stops it retrying forever?}}
Cite file paths and function names for every claim. If the answer
depends on code that is not included, name the missing file or module.
6. Find the root cause in a long log
For incident reviews. Asking for the first anomaly, not the loudest one, is the important part.
Below is the log from {{time_range:02:00 to 04:00 UTC}} during an incident where {{symptom:checkout requests timed out}}.
{{log}}
Based on the log above:
1. Give a timeline of the key events with timestamps.
2. Identify the first anomaly, which may come well before the errors.
3. Give the most likely root cause and the log lines that support it.
4. List what else could explain the symptoms and what would rule it out.
7. Check many cases against a long policy
Paste the full policy and a batch of cases. One long prompt beats checking each case in a new chat that has lost the policy.
<policy>
{{policy}}
</policy>
<cases>
{{cases}}
</cases>
For each case, decide whether it complies with the policy above. Return
a table: case ID, complies (yes, no, unclear), the policy section that
decides it, and a one-sentence reason. List the cases marked "unclear"
again at the end with the question that would resolve each one.
How do you get reliable JSON from DeepSeek?
In the API, set response_format to {'type': 'json_object'}, include the word json in the prompt, and show an example, as the JSON output guide describes. It also says to set max_tokens high enough that the JSON isn't cut off, and warns that the API may occasionally return empty content, in which case rewording the prompt helps.
8. Extract fields from messy text into JSON
For invoices, emails, job posts or support tickets. The example object does more than any description of the schema.
Extract the fields below from the text and return them as json.
Use null for anything not stated. Do not guess.
Example output:
{
"company": "Acme Ltd",
"contact_email": "ops@acme.example",
"amount": 1250.00,
"currency": "EUR",
"due_date": "2026-11-30"
}
Text:
{{text}}
9. Classify a batch into a JSON array
For sorting many short items in one call. A fixed label list and one object per item keep the output parseable.
Classify each item into exactly one of these labels:
{{labels:bug, feature_request, billing, account, other}}.
Return json with a single key "results" holding an array, one object
per item, in the same order, like this:
{
"results": [
{ "id": 1, "label": "billing", "confidence": "high" }
]
}
Confidence is high, medium or low.
Items:
{{items}}
Which DeepSeek prompts work in non-thinking mode?
Turn thinking off for tasks that need speed more than deliberation. In the API, set the thinking type to disabled in the request body. On the web, DeepSeek's announcements name Expert Mode and Instant Mode but don't spell out how they map to thinking, so check the chat box options. These are the tasks DeepSeek gives their own temperature recommendations for, so if you use the API, set the value noted under each.
10. Translate with a glossary
DeepSeek recommends temperature 1.3 for translation. The glossary keeps product terms consistent across a long document.
Translate the text below from {{source_language:English}} to {{target_language:Simplified Chinese}}.
Glossary (always use these translations):
{{glossary:Workspace = 工作区; Seat = 席位}}
Keep the formatting, links and code unchanged. Match the register of
the original. After the translation, list any phrase where you were
unsure of the meaning or where no good equivalent exists.
Text:
{{text}}
11. Clean and normalize records
DeepSeek recommends temperature 1.0 for data cleaning. Explicit rules stop the model from "fixing" data it should leave alone.
Clean the records below. Rules:
- Dates to {{date_format:YYYY-MM-DD}}
- Phone numbers to {{phone_format:E.164}}
- Trim whitespace and fix obvious capitalization in names
- Do not invent missing values; leave them empty
- Flag likely duplicates instead of merging them
Return the cleaned records in the same format as the input, followed by
a list of every change you were not certain about.
Records:
{{records}}
12. Several creative directions at once
DeepSeek recommends temperature 1.5 for creative writing. Asking for distinct directions uses that range instead of five near-copies.
Write {{count:4}} different versions of {{piece:a launch announcement}} for {{product}}.
Each version must take a different angle: for example one practical,
one story-led, one bold, one minimal. Label each with its angle in a
few words. Keep each under {{length:120 words}}. Audience:
{{audience:developers who are tired of marketing language}}.
13. Quick code fix with the smallest change
For small, well-understood bugs where thinking mode would be overkill. DeepSeek's recommended temperature for code is 0.0.
Fix the bug in this {{language:TypeScript}} function with the smallest
possible change.
Expected: {{expected}}
Actual: {{actual}}
{{code}}
Return the corrected function, then one sentence explaining the bug.
Do not refactor or rename anything else.
What should you know before pasting work data into DeepSeek?
DeepSeek's privacy policy, last updated 10 February 2026, says: "we directly collect, process and store your Personal Data in People's Republic of China." It lists user input among the data used to train and improve its models, and gives a right to opt out of training by emailing privacy@deepseek.com. Read it against your company's rules before pasting customer data or private code.
Because DeepSeek published the V4 weights openly, the other option is to run the model on infrastructure you control and send the same prompts there. For how temperature and tokens behave across models, see the LLM temperature guide and LLM tokens explained.
Keep your DeepSeek prompts in one place
Prompts like 4, 8 and 9 tend to become part of a workflow, so it pays to keep the working version somewhere you can find it. If you reuse one in an agent, the free SKILL.md generator turns it into a skill file in your browser.
To keep the prompts one shortcut away, save them in Kipt. Your prompts live as Markdown in a private GitHub repo you own, blanks like {{language:Python}} stay fillable, and Cmd+K finds any prompt by name or tag. Kipt is not affiliated with DeepSeek.