Editor's Note: DeepSeek R1 is exploding in popularity, and a lot of people who didn't use AI much are now trying to chat with big models. Today's post, from the perspective of an inference model, is about how to properly chat with the DeepSeek Founder Park is curating a closed-door discussion board on DeepSeek; see the end of this post for how to participate.
After DeepSeek exploded in popularity, I received over 50 requests for help from readers, almost all with similar struggles:
I really can't sit still when I see all this feedback.
I had planned to wait until after the New Year to write this advanced guide.
But I know all too well what people are missing out on.
It's like watching a bunch of guys take an East Wind missile to a fish fryer and can't get enough of a splash.
So, I decided to go ahead and release this DeepSeek Cheat Sheet.
Why does it have to be released before Spring?
Because when you watch the spring show and then study it.
The first people to capitalize on the true value of DeepSeek have long since grabbed all the dividends of your industry.
Yes, the free Deepseek has caught up to the $200 a month chatGPT-o1 and even partially surpassed it.
Like the moment in 2011 when the first flight of the J-20 tore through the night sky--
We finally don't have to be on our knees with AI.
Note: All the tips in this article are from real cases and all the cue words have been verified over and over again.
01 The most important secret:Throw away your cue word template
If you're still using "pro tips" and "templates", you're going in the wrong direction.
DeepSeek doesn't eat that at all.
Why?
Because it is at its core a reasoning macromodel, not a directive macromodel.
It's like two interns:
- A little nerd who needs you to organize the steps of a task in minute detail. (directive type)
- A little wit who can think for himself what to do if you explain the purpose. (reasoning type)
Let me illustrate with a real-life example that
A real-world test by one of our community's operations students to conduct a new energy industry analysis for use in preparing for negotiations with BYD suppliers.
The traditional way:
You are asked to play the role of a new energy industry analyst and follow the steps below:
1. Market size
2. Competitive landscape
3. Technical tracks
4. Future trends
Requirements: 800 words per section, citing authoritative data...
Result: get a dry report with a glance at AI.
A glance at the AI is nothing but correct gross.
The right way:
I'm negotiating with a BYD supplier next week and know nothing about power cells. Help me explain in the most layman's terms:
1. What are their technological advantages
2. How much is likely to be asked for
3. What is the terminology that can be used in negotiations?
The point is to make me listen and act like an insider.
RESULT: DeepSeek gives a direct grounding analysis with negotiation talk. 
I can't believe you've got words in your mouth.
That's the big difference:
DeepSeek doesn't require you to write "professional tips".
It requires real scenarios and specific needs.
Sending you a generic formula:
I want xx, for xx, hoping for xx results, but worried about xx issues...
It's like you're talking to an intelligent subordinate:
- Don't say, "Please write your weekly report in accordance with the STAR Code."
- And to say:
I'm going to write a weekly report, the boss is going to read it on Monday, and I hope the focus will be on xxx, the focus is so that our department can achieve the effect of pretending in front of the boss, and press the R&D department next door, but I'm worried that the R&D is questioning the fact that we don't have enough detail in the writing of the product documentation ......"
02 Most underrated feature:Make it "human."
Many people complain that DeepSeek's responses are too abstract, like they're reading from a book.
Like this one. Don't believe it.
But what you may not realize is that all it takes is one simple cue word to completely change the problem.
The magic cue word is:
Speak human.
That's right, just those three words.
My trainee didn't believe me on the first try, but it turned out...
Original Answer:
Abstract. Fucking abstract.
Add the answer after "Speak human":
afterword
Instant grounding, right?
Because deepseek is sensitive to the word "human".
Of course, sometimes those three words aren't enough, and you can use this exhaustive version of the prompt word and just copy it over:
Please use the following normative output: 1. language plain and straightforward, avoiding abstract metaphors; 2. use daily scenarios to assist in the illustration of the case; 3. give priority to the choice of specific terms instead of abstract concepts; 4. keep paragraphs concise (no more than 5 lines); 5. technical expressions need to be accompanied by common sense explanations; 6. prohibit the use of literary rhetoric; 7. focus on the information in front of the; 8. complex content is illustrated by sub-points; 9. keep the spoken language, but do not overly Simplify specialized content; 10. Ensure the accuracy of the information under the premise of giving priority to the selection of popularly recognized vocabulary].
03 The most powerful skill: deep thinking
Here's what I have to say.
A free homegrown AI is making GPT-o1, which pays $200 a month, sit up and take notice.
Why?
Because DeepSeek thinks in a way that is smarter than GPT-o1.
Let me contrast this with a real-life example:
GPT-o1's answer:
DeepSeek's answer:
Reasoning, not linear listing
That's the big difference:
- GPT-o1 linear listing, like an advanced documentation tool
- DeepSeek Deep Thinking, Like a Thinking Partner
Free DeepSeek, which directly puts the cowhorses of AI companies across Silicon Valley into overtime, all night long.
Luckily, they don't have to celebrate the Spring Festival.
But recently, I have noticed a phenomenon:
DeepSeek has apparently adjusted its response strategy due to the surge in users:
- Thinking time reduced from 20 seconds to 5 seconds
- Significant decrease in depth of response
- Limited ability to reflect
The math is tight. We'll only give you five seconds.
It's an understandable temporary measure; after all, arithmetic is all about burning money.
But how do we, the users, continue to stimulate its ability to think deeply?
I've compiled three core cue words, called the Deep Thinking Trio for the sake of pretentiousness:
- Please conduct at least 10 rounds of critical thinking along with your thoughtful analysis, and be sure to be thorough!
- Please consider your answer at least 10 rounds from the opposite side at the same time as your thinking and analyzing process, and be sure to be thorough!
- Please review your responses for at least 10 rounds at the same time as your thoughtful analysis, and be sure to be thorough!
In this way, deep thinking will be restored from 5 seconds to about 20 seconds.
For the italicized portions, feel free to replace them with forms you are comfortable with, or combine and stack them.
butReflection at the coreThe
Deep thinking goes back to 20 seconds.
04 The most powerful style converter
Yesterday, I used DeepSeek to write a Hanfu.
Praise for Wang Xing's affectionate and intelligent girlfriend. Wang Xing is the actor who was sold to North Burma a while back.
I had a great time reading it myself.
There's really no one else for this use of diction, this ekphrasis.
After posting it to the group, it blew up three language teachers...
the most powerful weapon in the arsenal
Then there are friends who write words that even professional screenwriters say are old-fashioned.
It's very flavorful. 
Professional screenwriters say it's great.
In fact, it's DeepSeek's fourth secret weapon: the style converter.
The cue words are simple, even if you just dictate them:
Write a xx style of writing about xxxxxx in imitation of xxxxxx.
But be aware of two limitations:
On the one hand, it is more suitable for imitating classic authors because of sufficient training data and distinct stylistic features.
On the other hand, don't expect a 100% restoration, the 80% resemblance is already impressive, the point is to be able to capture the charisma.
I think deepseek is definitely trained in a large corpus of literature, especially classical Chinese literature.
Lu Xun and Mo Yan, for example, are exquisite.
Imitation of Lu Xun's style
Remember that universal formula:
"
I want xx, for xx, hoping for xx results, but worried about xx issues...
Example:
"
I'm going to write a fugue about praising Wang Xing's girlfriend's wit and bravery with heart to show off on top of Little Red Book, hopefully focusing on imitating Wang Bo's piece of writing, focusing on letting my own literary skills pretend to get a hundred million likes on Little Red Book, but worrying that others won't understand it's too obscure ......"
A similar stylistic migration can be placed on contemporary writers, such as our parody of Mr. Run Liu's New Year's Eve speech:
Feel free to imitate the prompt words
Note that you have to provide the original content first (as detailed as possible, usually no less than 8,000 words), and then just ask them to parody it directly.
Of course, it's best to use the universal formula for better results.
05 Use of prohibited areas: what not to use it
Having said all these advantages, its limitations must also be mentioned.
DeepSeek is not recommended for the following scenarios:
1. Long-form writing
Articles over 4000 words are prone to logical breaks and Claude 200k is recommended.
Because deepseek defaults to 64k, long texts are not enough.
Long text is really not enough at the moment
2. Sensitive content
After all, it's a homegrown AI with a built-in audit scale that's deranged.
Any time you reply like this, it's triggering a review
Many times you don't know which sentence triggered the review.
How can this situation be resolved?
Because deepseek is post audit, there are three options as follows:
1. click Modify at your question and submit it a couple times, there is always a time when it doesn't trigger a review.
Click Edit at the question
2. When generating your answer, furiously tap the Copy Answer button. This ensures that when the review is triggered, you have the content of the previous answer on top of your clipboard!
Tap the answer button
3. Once and for all, change the imperial three. (gpt, claude, gemini)
3. Personal style writing
Without going into detail, this is a reasoning model, suitable for problem solving and imitation.
But it's hard to ensure that you get the style writing effect you want by controlling it precisely.
This is not really a disadvantage of deepseek, only a feature.
I'll address the specifics in my next post.
06 How will it change our AI era?
After a few days of intensive testing, I'm becoming more and more convinced:
DeepSeek represents the future direction of AI -- to understand human language better and to think better.
You don't need to learn its language, it's learning to understand yours.
What does that mean?
The barrier to using AI is rapidly decreasing.
In the future, we won't need to memorize cue word templates, learn specific instructions, and study various parameters.
Just, say what you want, tell it specific scenarios, and speak human.
Because DeepSeek took out the most anti-human setting of AI in 23 years
--Let humans learn machine language (cue words).
Actually, the original intent of the Wunderkind Cue Word Engineering class was a good one
It's like a key turning point in the evolutionary history of cell phones:
- Nokia era: read the manual, learn the key combinations, swipe the phone
- The iPhone era: three-year-olds can play with it
My friends in the investment community put it more bluntly:
Vloggers who are still teaching people to write prompts in 2024 are charging an IQ tax.
Anyway, there's a shitload of cue words to learn!
So, that's why I'm going to insist on posting this article before the New Year.
Deepseek is the paradigm that represents the next generation of AI usage, resigned to the new!
For the record, deepseek didn't pay me a dime, but I'm just going to blow up.
07 Domestic AI's J-20 Moment
We've waited too long for this day.
When I used the GPT-4 for the first time, on March 15, 2023, my hands were shaking.
On the one hand because it's powerful.
On the other hand, it's also because I know:
At this level of AI, we may be chasing a decade.
Until deepseek released the R1, I continued to give it a try with skepticism and no hope at all.
But seeing the depth of thought that went into it gave me an answer that was completely unmatched, if not partially surpassed, by the Mikado (GPT, claude, gemini).
My hands shook again.
I know, times have changed.
DeepSeek showed me:
- It understands the human condition behind "speaking human".
- It understands the rules of socialization in the "fake effect".
- It can write about the absurdity of the fraud in northern Burma with the Tengwang Pavilion Preface
While Silicon Valley is still teaching users how to "domesticate AI," DeepSeek is doing something much bigger:
Teach AI to understand the human mind.
Crucially, DeepSeek is still open source, and this is the first time the open source world of AI has caught up with the closed source world in a bright way.
Like Yang Wei, the chief designer of J-20, said:
"We're no longer playing catch-up, we're defining the new battlefield."
This is perhaps the best New Year's gift of all:
For the first time in the race for AI, we stand at the same dawn as the world.
English 
简体中文 
日本語 
Deutsch (Sie) 
Português do Brasil