DeepSeek has become a phenomenal AI product with greater speed than ChatGPT and is one of the most discussed topics this Spring Festival. For the majority of people, the most necessary thing to do at present is to start using it more. Perhaps 99% of people have not yet used this AI correctly. This article is a DeepSeek usage guide summarized from real-life use cases. If you have also discovered more useful tips in your use, welcome to share them in the comments section.
After DeepSeek’s explosive popularity, I received requests for help from over 50 readers, almost all with similar troubles: hence, I decided to publish this ‘DeepSeek Strategy Guide’. The first batch of people who truly utilized the value of DeepSeek have already seized the dividends in your industry. Yes, the free DeepSeek has caught up with the $200 per month ChatGPT-o1, and even surpassed it in some aspects. Just like the moment when the J-20’s maiden flight tore through the night sky in 2011 – we no longer have to kneel to use AI. Note: All tips in this article come from real cases, and all prompt words have been verified repeatedly.
There are even scripts. This is the biggest difference: DeepSeek doesn’t require you to write “professional prompts”. What it needs are real scenarios and specific requirements. Here’s a general formula for you: I want xx, for xx to use, hoping to achieve xx effect, but worried about xx problem… Just like talking to a smart subordinate: Instead of saying “Write the weekly report according to the STAR rule”, say: I need to write a weekly report.
The boss will review it on Monday. I hope to focus on xxx. The key is to make our department stand out in front of the boss and outshine the R & D department next door. However, I’m worried that the R & D department will question that our product documentation is not detailed enough… The most underestimated function: Make it “speak human language”. Many people complain that DeepSeek’s responses are too abstract, like reading a天书 (indecipherable text). For example, don’t believe this. But what you may not know is that with just a simple prompt, this problem can be completely solved. This magical prompt is: Speak human language. That’s right, just these three words. My students didn’t believe it when they tried it for the first time. As a result… Original answer: Abstract, extremely abstract. Answer after adding “Speak human language”: It immediately becomes more down-to-earth, right? Because DeepSeek is very sensitive to the term “Speak human language”. Of course, sometimes these three words are not enough. You can also use this detailed prompt. Just copy it directly: [Please output according to the following specifications: 1. Use plain and straightforward language, avoiding abstract metaphors; 2. Use daily scenario-based examples for illustration; 3. Prefer specific nouns to replace abstract concepts; 4. Keep paragraphs concise (no more than 5 lines); 5. Technical expressions should be accompanied by popular explanations; 6. Literary rhetoric is prohibited; 7. Place key information at the beginning; 8. Present complex content in points; 9. Keep it colloquial but don’t oversimplify professional content; 10. Prioritize commonly recognized vocabulary on the premise of ensuring information accuracy] The most powerful skill: Deep thinking. This is something I have to say: A free domestic AI is making o1, which costs $200 per month, feel uneasy. Why? Because DeepSeek’s way of thinking is more intelligent than o1’s. Let me compare with a real case: o1’s answer: The picture is from Teacher @D&roi. DeepSeek’s answer: Reasoning, rather than listing linearly. This is the biggest difference: GPT-o1 lists linearly, like a sophisticated document tool. DeepSeek thinks deeply, like a thinking partner. The free DeepSeek has directly made the employees of AI companies in Silicon Valley work overtime overnight. Fortunately, they don’t celebrate the Spring Festival.Recently, I’ve noticed a phenomenon: Due to a sharp increase in users, DeepSeek has significantly adjusted its response strategy. The thinking time has been reduced from 20 seconds to 5 seconds, the depth of answers has decreased significantly, and the reflection ability is limited. Given the tight computing power, setting only 5 seconds is an understandable temporary measure, after all, computing power incurs costs. But for us users, how can we continue to stimulate its deep thinking ability?
I’ve sorted out three core prompt words. For the sake of showing off, let’s call them the ‘trilogy of deep thinking’: – Conduct critical thinking for at least 10 rounds simultaneously during your thinking and analysis process. Be sure to be thorough. – Consider the opposite side of your answer for at least 10 rounds simultaneously during your thinking and analysis process. Be sure to be thorough. – Conduct a review of your answer for at least 10 rounds simultaneously during your thinking and analysis process. Be sure to be thorough. In this way, deep thinking will be restored from 5 seconds to about 20 seconds. The italicized parts can be freely replaced with forms you’re good at, or combined and superimposed. The core is reflection. Deep thinking returns to 20 seconds. The most powerful writing style converter. Yesterday, I used DeepSeek to write a Hanfu. It praised the affectionate, brave and resourceful girlfriend of Wang Xing. Wang Xing is the actor who was sold to northern Myanmar some time ago. I really enjoyed reading it myself. The use of allusions and parallel prose are truly remarkable. After posting it in the group, it directly attracted three Chinese teachers… It’s really a ‘group-stirring’ artifact. Some friends have written texts that even professional screenwriters say are sophisticated. They have a unique charm. Professional screenwriters all say it’s great. In fact, this is precisely the fourth secret weapon of DeepSeek: the writing style converter. The prompt words are very simple. You can even just state them orally: Imitate the writing style of xxx and write a xx-style article about xxxxx. But note two limitations: On the one hand, it is more suitable for imitating classic writers because of sufficient training data and obvious style characteristics. On the other hand, don’t expect a 100% restoration. An 80% similarity is already remarkable. The key is to capture the charm. I think DeepSeek must have trained on a large amount of literary corpora, especially classical Chinese literature. For example, imitations of Lu Xun and Mo Yan are very vivid. Remember the universal formula for imitating Lu Xun’s style: I want xx, to be used for xx, hoping to achieve xx effect, but worried about xx problem… For example: I want to write a rhapsody praising Wang Xing’s girlfriend for her wisdom, courage and dedication, to show off my skills on Xiaohongshu. I hope to focus on imitating Wang Bo’s parallel prose, with the key being to show off my literary talent and get 100 million likes on Xiaohongshu, but I’m worried that others may find it too obscure… Similar style transfers can also be applied to contemporary writers. For example, when we imitate Teacher Liu Run’s New Year’s speech: The prompt words for imitation can be whatever you like. Note that you need to first provide the original content (try to be as detailed as possible, generally not less than 8000 words), and then directly ask it to imitate.Emulating the effectiveness of Liu Run’s approach, for better results, it is best to employ the versatile formula. However, there are certain situations where it should not be used. After discussing its advantages, we must also address its limitations. The following scenarios are not recommended for using DeepSeek: 1. Long text writing exceeding 4000 words tends to result in logical discontinuity; it is suggested to use Claude200k instead.
Since DeepSeek defaults to 64k, it is insufficient for long texts. Currently, it is indeed inadequate for long texts. 2. Special content As a domestic AI, it has. Whenever such a response is triggered, you may not know which sentence has activated it. How to resolve this situation? Since DeepSeek is a post-moderation AI, there are three solutions: 1. Click to edit in your question area and resubmit several times; there will be at least one instance that does not trigger moderation. Click on the question area to edit. 2. While generating a response, repeatedly click the copy response button. This ensures that when moderation is triggered, your clipboard contains the content of the previous response. Click the answer button. 3. A one-time solution is to switch to the top three AI platforms (GPT, Claude, Gemini). 3. Personal style writing This aspect will not be elaborated upon, as it is a reasoning model suitable for problem-solving and imitation. However, it is difficult to ensure the desired style of writing through precise control. This is not actually a disadvantage of DeepSeek, but rather a characteristic. I will discuss this in more detail in my next article.How will it change our AI era? After several days of intensive testing, I am increasingly convinced that DeepSeek represents the future direction of AI — more human-like understanding and thinking.
You don’t need to learn its language; it is learning to understand yours. What does this mean? The barrier to using AI is rapidly decreasing. In the future, we won’t need to memorize prompt templates, learn specific commands, or study various parameters. Simply, clarify what you want, inform it of the specific scenario, and speak in human language. Because DeepSeek has eliminated the most anti-human aspect of AI for 23 years — making humans learn machine language (prompts). Actually, Andrew Ng’s initial intention with prompt engineering was good. This is like a key turning point in the evolution of mobile phones: Nokia era: Read manuals, learn combination keys, flash the phone. iPhone era: Even a three-year-old can play with it. My investment circle friends put it more directly: In 2024, influencers who still teach people to write prompts are collecting an intelligence tax. In short, there’s no need to learn prompts anymore! Therefore, this is also why I insist on publishing this article before the New Year.DeepSeek represents the new paradigm of AI usage, bidding farewell to the old and welcoming the new! We have been waiting for too long for the moment of domestic AI like the J-20 stealth fighter. On March 15, 2023, when I first used GPT-4, my hands were trembling.
On one hand, because of its power. On the other hand, because I knew: we might have to catch up for a decade with AI at this level. Until DeepSeek released the R1, I still tried it with a skeptical mindset, without any hope, but seeing its deep thinking, the answers it provided were not inferior to, and even surpassed, those of the top three (GPT, Claude, Gemini) in some aspects. My hands trembled again. I knew the era had changed. DeepSeek showed me: it understands the human relationships behind ‘speaking human language’, it comprehends the social rules within ‘showing off’, and it can use ‘Tengwang Pavilion Preface’ to write about the absurdity of fraud in Myanmar. While Silicon Valley is still teaching users how to ‘tame AI’, DeepSeek is doing something greater: teaching AI to understand human thinking. Crucially, DeepSeek is open-source, marking the first time an open-source AI has openly caught up with closed-source AI. As Yang Wei, the chief designer of the J-20, said: we are no longer catching up, we are defining new battlefields. This might be the best New Year’s gift: in the AI competition, for the first time, we stand at the same dawn with the world.
