How do you write a good DeepSeek prompt?

For the regular chat model, write it the way you would for ChatGPT: the task, the facts it cannot guess, the rules and the shape of the answer. For the reasoning mode, cut the scaffolding instead. State the goal, list the facts and constraints, say how the final answer should look, and leave out personas, worked examples and instructions about how to think. DeepSeek reported that few-shot examples made its R1 reasoning model worse, so describe the problem directly.

Should I use a system prompt with DeepSeek?

With the chat model, a system prompt works as it does elsewhere. For the reasoning model, DeepSeek originally advised putting every instruction in the user message and no system prompt at all; a later R1 update said system prompts are supported. The safe habit is to keep anything that matters in the user message, where every version reads it.

Is Grok better than ChatGPT for current events?

Grok can read posts on X, which makes it quick at picking up what people are saying right now. Fast is not the same as accurate: in the Tow Center audit of AI search tools in February 2025, Grok 3 answered 94% of citation queries incorrectly. Ask Grok for links, give it a time window, and open the links before you rely on them.

Can I use the same prompt for ChatGPT, DeepSeek and Grok?

A clear prompt with the task, context, rules and output format works in all three. The changes are small but they matter: strip the step-by-step scaffolding for a reasoning model, and spell out tone and allowed sources for Grok. If you are comparing the models, send all of them the same neutral prompt so you measure the model rather than the prompt.

DeepSeek and Grok Prompts: What Changes from ChatGPT

Most of what you know about prompting ChatGPT carries over to DeepSeek and Grok. All three read plain language, and a prompt that states the task, the facts, the rules and the shape of the answer gets a usable result from any of them. The differences sit in two places: DeepSeek’s reasoning mode wants less instruction about how to think and more about what a good answer looks like, and Grok needs you to set the voice and the sources yourself, because its defaults lean further from neutral business prose than ChatGPT’s do.

Below: what to keep, what to change and what to drop, with before and after prompts. Model versions change every few months, so treat each point as a hypothesis to test on your own work and re-check it after the next release.

The quick answer

Prompt habit from ChatGPT	DeepSeek chat model	DeepSeek reasoning mode	Grok
One-line role (“You are an editor”)	Keep	Optional; facts and checks matter more	Keep
”Think step by step”	Keep if it helps	Drop; it already reasons before answering	Keep if it helps; drop for its reasoning modes
Worked examples of the answer you want	Keep	Drop; describe the format in words instead	Keep
Long system prompt or custom instructions	Works	Move anything important into the user message	Works
No tone instruction, trusting the default	Usually fine	Usually fine	Add one; the default is more casual
”Search the web” with no limits	Depends on the app	Depends on the app	Name the sources, the time window and ask for links
Explicit output format	Keep	Keep, and put it last	Keep

If you remember one line: tell a reasoning model what done looks like, and tell Grok who it should sound like and where it may look.

Start from a prompt that works anywhere

Before tuning for a model, get the neutral version right: the task, the facts the model cannot guess, the rules (length, tone, what to do when information is missing), the shape of the answer, and, when it helps, a one-line role. Square-bracket placeholders such as [DEADLINE] mark facts you will fill in, so the model never invents them.

Use this neutral version if you are deciding between models. A prompt tuned for one model carries small fixes for its habits, and pasting it into another measures the fit of the prompt rather than the ability of the model. The full method is in how to compare AI models side by side.

DeepSeek’s chat model: mostly portable

DeepSeek’s regular chat model behaves like the assistants you already know. A prompt written for ChatGPT will usually work unchanged, system prompts work, and examples of the output you want help the way they do elsewhere. Two adjustments are worth making.

Say which language to answer in. DeepSeek has said its reasoning model is optimised for Chinese and English and can mix languages when asked in others. If your input mixes languages, or you paste a source in one language and want the answer in another, add a line such as “Answer in English, even where the source is in Spanish.” It costs nothing and removes a guess.

Spend the savings on context. DeepSeek models are among the cheapest in most model pickers, so it is reasonable to paste the whole document instead of a summary, or to run the prompt twice and keep the better answer.

DeepSeek’s reasoning mode: say what, not how

The reasoning mode (the R1 line, and the thinking mode in later releases; in DeepSeek’s own app it has sat behind a DeepThink button) works through a problem at length before it writes the answer. That changes what a good prompt looks like, and DeepSeek has said so directly. In the R1 paper, the team reported that few-shot prompting consistently made the model worse and recommended describing the problem directly and specifying the output format, with no examples. Their usage notes for the first R1 release also advised putting all instructions in the user message rather than a system prompt; a later update said system prompts are supported.

In practice that means four changes from a ChatGPT habit:

Drop the persona and the thinking instructions. “You are a world-class analyst, think carefully step by step” adds little to a model that reasons by default. One documented exception: for maths problems, DeepSeek suggested asking it to reason step by step and put the final answer in a box, so the result is easy to find.
Drop the worked examples. If you want a particular format, describe it in words.
Give it facts and checks. A reasoning model gets more out of the numbers it needs and a test the answer must pass than out of advice on method.
Put the output format last. The model writes a long chain of reasoning before the answer; a format line at the end of the prompt is the last thing it reads before it starts.

Here is a ChatGPT-style prompt:

You are a senior SaaS pricing consultant with 20 years of experience. Think step by step and be thorough.

Here are two examples of the kind of analysis I like:
[EXAMPLE 1]
[EXAMPLE 2]

Should we add annual billing?

And the same request rewritten for the reasoning mode:

Decide whether we should offer annual billing at a 15% discount alongside our monthly plan.

Facts:
- [NUMBER] monthly subscribers at [PRICE] a month
- Monthly churn: [CHURN RATE]
- Cash in the bank covers [MONTHS] months of costs

Constraints:
- Over the next 12 months, cash flow matters more than total revenue
- Use only the facts above; label anything else you assume as an assumption

Before you answer, check that the recommendation still holds if churn is half or double the figure above.

Output: a one-line recommendation, the churn rate at which the two options break even, and the three assumptions that would change the answer.

The second prompt is longer, but every added line is a fact, a limit or a test. Nothing in it tells the model how to think.

If you run the open-weight model yourself, or through a host that exposes sampling settings, DeepSeek’s R1 notes recommended a temperature between 0.5 and 0.7, with 0.6 as the default to try, to avoid repetitive or incoherent output. DeepSeek’s own app and API, and most chat apps, set this for you.

Grok: set the voice and the sources yourself

xAI introduced Grok as an assistant with “a bit of wit” and “a rebellious streak”, built to answer questions other assistants decline, with live access to posts on X. All three traits show up in prompts. ChatGPT’s default voice is polished and careful; Grok’s leans more casual and more willing to take a side. If you want something specific, ask for it.

Name the voice and the things to leave out. A one-line task that ChatGPT turns into a tidy corporate draft can come back from Grok with a joke in the opening line. Before:

Write a LinkedIn post announcing our new pricing.

After:

Write a LinkedIn post announcing our new pricing.

Voice: plain and neutral, like a product changelog. No jokes, no slang, no exclamation marks, no hashtags.
Facts: the old price is [OLD PRICE], the new price is [NEW PRICE], it starts on [DATE], and existing customers keep [WHAT THEY KEEP].
Length: under 120 words.
Do not mention competitors or speculate about why prices changed.

The voice line does most of the work here. It is worth adding for ChatGPT too, but with Grok it is the difference between a draft you can post and one you rewrite.

Put your own limits on live information. Grok’s access to X makes it fast at picking up what people are saying right now, and X posts are a noisy source. Independent audits of AI search have not been kind to it either: in the Tow Center test from February 2025, Grok 3 answered 94% of queries incorrectly and 154 of its 200 citations were broken or fabricated links. Newer versions may do better, but the findings are a reason to ask for links you can open, not a reason to trust the summary. The details are in can you trust AI search citations. Before:

What are people saying about [PRODUCT]?

After:

Summarise reactions on X and in news coverage to [PRODUCT] from the last 48 hours.

For each theme: one line describing it, whether it is common, occasional or rare in what you found, and links to two posts or articles that show it.
Prefer posts from accounts that name a person or an organisation.
If there is not enough material in that window, say so instead of widening it.
Mark any claim you could not trace to a link as unverified.

The time window, the source list and the instruction for the empty case are the three lines that do the most good. Without them, a thin result tends to get padded with older or weaker material.

Add the guardrails you used to get for free. Because Grok declines fewer requests by default, the limits that ChatGPT applies on its own (not naming private individuals, not guessing at someone’s motives, flagging speculation) are yours to write when they matter. xAI publishes the system prompts behind the Grok app on GitHub; reading them once shows you which instructions your prompt is sitting next to.

Grok’s reasoning modes follow the same logic as DeepSeek’s: state the goal, the facts and the test, and leave the method to the model.

Two ChatGPT habits that do not travel

Leaning on custom instructions and memory. If your ChatGPT setup knows your job, your audience and your preferred format, your prompts have probably got shorter without you noticing. None of that comes with you to DeepSeek or Grok. Paste the missing context into the prompt, or save it as a reusable block you add at the top.

Assuming the refusal line is in the same place. ChatGPT, DeepSeek and Grok draw it in different places, in both directions. DeepSeek’s own app declines some politically sensitive topics that the others will discuss; Grok answers some requests that ChatGPT declines. If your prompt depends on the model refusing or not refusing something, test that directly rather than assuming.

Check the advice on your own prompts

Take two or three prompts from last week, write the neutral version and the tuned version for each model, and read the answers side by side. If the tuned version does not beat the neutral one on your work, keep the neutral one.

If you would rather not rewrite each prompt by hand, a DeepSeek prompt generator turns a one-line idea into a goal, constraints and checks layout suited to the reasoning mode, and a Grok prompt generator adds the voice, source and freshness lines Grok needs. Edit the result with your real facts before you send it.

To run the comparison without three accounts, AIHub has DeepSeek V4 Pro, DeepSeek V4.1 Flash, Grok 4.7 and Grok 4.3 in the same model picker as GPT-6 Sol, Claude and Gemini, on one bill. Pick one model, add a second, and the same prompt goes to both with the answers next to each other, so you can see whether the prompt or the model made the difference.