The AI Writing Guide 1.0: The Vessel of Intelligence Outweighs Intelligence Itself

Over the past week, I gave two internal presentations at Huxiu about AI writing. The thinking behind these talks is something I believe is worth sharing broadly. As Wenfeng Liang, founder of DeepSeek, put it: in AI, there are no real technical moats — organization and culture are the core competitive advantages. If even a large model company says this, why can't media thinking be shared?

On AI writing, I have one key argument: AI can completely replace all business media writing. The output of outlets like 36Kr, Huxiu, and LatePost can essentially all be handled by AI. By using AI for writing, we can easily improve text workflow efficiency by two to three times. And we need to believe this: AI writes better. It's more logical than humans. It's faster than humans.

But this requires us to rethink what AI actually is. At its core, AI is a probabilistic model. It excels at execution, writes beautifully, and reasons well. But AI has no self-awareness. It cannot generate truly original ideas. It merely predicts the probabilistic relationships between words — it doesn't possess human creative thinking.

For example, in LatePost's interview with He Xiaopeng, the things He said — no AI could have originally produced those in their entirety. AI might phrase certain words more elegantly than He's spoken remarks. But the crucial point is: humans have original ideas; AI doesn't. That's our core competitive advantage as people.

Humans Are the Method; AI Is the Tool

Our relationship with AI needs to shift. There's a common misconception — you see certain old-school bosses who, after using AI, send messages to their staff groups saying, "We can cut 80% of employees; the remaining 20% just need to know how to use AI." This is wrong.

Using AI is deceptively simple. Chat with DeepSeek for a few minutes and you get some pretty decent responses. But this doesn't actually improve work efficiency. To make AI perform at maximum capacity — to replace writing and boost text workflow efficiency by two to three times — requires some engineering techniques for writing. Using AI isn't hard; asking it a few questions is easy. What's truly difficult is integrating AI into your workflow, making AI's intelligence flow like water to where it needs to go.

So returning to the roles of humans and AI, our most important shift is in mindset. Humans shouldn't compete with AI on writing speed, prose quality, or logic. Humans are the method; humans are the purpose; humans are the leaders. AI is merely an execution tool. We should treat AI as an efficiency-enhancing tool, while humans provide original ideas, creativity, and decision-making. Just as after the Industrial Revolution, there was no point in humans competing with machines on strength or precision — instead, we should command machines, design machines, and let machines do the work to improve production efficiency.

The same applies to content creators. AI hasn't eliminated content creators at all; rather, it's an enormous help. A junior writer can now operate like a senior editor — no longer needing to execute personally, just communicating ideas to AI and letting it help break down concepts and handle the writing. Humans should invest energy into more creative, original thinking rather than wasting it on concrete execution work.

So how does AI integrate into our workflow? Some very common approaches are: dumping a document into the DeepSeek app and asking it to rewrite it into an article; or sending it a message saying "I need to write a proposal, help me draft it." These approaches don't work.

Because, returning to our starting point: AI is fundamentally a probabilistic model. It doesn't possess original ideas; it doesn't truly understand what an article means. It just predicts between words. If you feed it an article, it can indeed understand the structure well, but it cannot truly grasp your intended meaning. It will just fill the structure with plausible-sounding but ultimately unusable content that looks sophisticated but doesn't actually work.

Therefore, all our work in integrating AI into workflow should revolve around one core principle: how to avoid AI hallucinations, and how to make AI accurately execute our ideas.

Outline First, Full Text Second

I've summarized the core principle of AI writing in eight characters: Outline first, full text second.

It sounds simple, but it means that when using AI for writing, you need your own core perspective. For topics you understand, you can directly dictate your thoughts and have AI generate an outline. For unfamiliar topics — say, writing about semicon chips when you've never worked in semiconductors — you need to feed it substantial material, keep questioning it until you're certain you understand the key concepts and can explain the topic using them, and only then have AI write your core ideas into an outline.

After generating the outline, you need to specifically check whether AI accurately executed your ideas, whether there are deviations. For instance, if a case study in one section needs to move to another, you need to give very specific instructions to get AI to align with your thinking. Only when AI produces an outline that accurately reflects your ideas should you have it write the full text based on that outline.

The second key point is segmented writing. Because AI gets lazy. Among current major models, only Google's Gemini 2.0 thinking reasoning version can write longer pieces (over 2,000 words). Other AIs including Claude 3.5 Sonnet and DeepSeek R1 suffer from severe hallucinations when writing long articles. So we need a modular writing approach, breaking long articles into segments.

Take a typical 3,000-word business media article. With an accurate outline in hand, you have AI write the beginning, middle, and end separately. After each section, you give specific instructions. For example, after writing the opening, if you feel it's not engaging enough, you can tell AI that certain material would work better at the start; or if the tone feels off, tell AI the tone you want and have it rewrite.

Here's a crucial realization: You cannot use AI to write something beyond your own cognitive scope.

As Wittgenstein said, "The limits of my language mean the limits of my world" — this is exactly what he meant. The prerequisite for articulating accurate demands is having a relatively clear idea of what you want to do. If my skill level is poor, there's no way I can get AI to write something at a high level for me.

The third key point is verifying AI's errors. There's a very direct method: throw the AI-written article back at AI and have it verify all specific data and core statements against your provided materials, finding original sources one-to-one. You can have it check once or twice until you're certain all statements are accurate.

Regarding specific AI writing techniques, there are many prompt templates online, but these don't matter much. Because with each new model release, the specific template requirements change. For example, with the rise of reasoning models like DeepSeek, those previously popular highly detailed role-setting and step-by-step prompt templates actually conflict with reasoning models' characteristics and weaken their capabilities.

What matters is integrating AI into your workflow so that the workflow naturally becomes your prompt. Once you've streamlined your workflow, you can have AI directly generate a prompt for this workflow — like genetic code that can iterate continuously. You can even build a prompt library, using different prompts for different types of work.

On model selection specifically, I don't recommend using DeepSeek alone. It's currently unstable, with too much hallucination, not really suitable for work. For daily use, I more recommend Claude, ChatGPT, or Google Gemini — especially Gemini 2.0 thinking, which works very well, comparable to DeepSeek R1, just less hyped.

There are also some interesting open-source projects now that combine multiple large models' capabilities. For example, DeepClaude combines DeepSeek's reasoning capability with Claude's execution capability — having DeepSeek do the reasoning and Claude handle the writing.

This is an open-source project by a friend of mine (github.com/ErlichLiu/DeepClaude), which I consider currently the strongest AI writing solution. If you don't want to deploy this open-source solution yourself, you can also purchase DeepClaude API access from him.

The Container of Intelligence Is Greater Than Intelligence Itself

In the AI era, content creators need to rethink their positioning and value.

My first suggestion: strategically despise AI, tactically respect AI. At the strategic level, you need to understand and recognize AI's limitations. You need to believe that we humans have original thinking, that we are the method, that we are the purpose itself. At the tactical level, we need to respect AI, because its execution capability is genuinely powerful. Its prose quality, logic, and writing efficiency far exceed human capabilities.

AI will intensify the Matthew effect in content creation. A 100-point creator can use AI to mass-produce 90-point content, while a 50-point creator can only use AI to mass-produce 60-point content. So influence gaps in the content industry will be further amplified. What we need to do is seize this time window, quickly master AI, and use it to massively replicate our current cognitive level. In the long term, what humans truly need to do is raise our cognitive level so we can continuously generate new ideas.

After repeated discussions with friends, I've reached a consensus: writing itself has now been replaced by AI. Content creators' core competitive advantages are topic selection, emotional resonance, and exclusive case studies.

When everyone can use AI to mass-produce content, your ability to pull off something stunning that catches everyone's eye, to write an article with intense emotion that moves readers, or to interview important figures and obtain exclusive information — these are your core competitive advantages.

There's a metaphor that explains the current situation: it's like a marsh versus the Three Gorges Dam. Our original work was like a marsh — a complex ecosystem that might produce some rare, precious fish. But now AI is like the Three Gorges Dam. Do you connect your marsh to the Three Gorges project? Let the surging Yangtze waters transform your marsh, redesign your water system and ecosystem?

Of course, you can absolutely stay in your marsh, pursuing some rare content form, producing genuine literary work. But now that the Three Gorges Dam has arrived, we have the opportunity to connect our marsh to it. Use AI's computing power, like Yangtze waters, to flush through our workflow and restructure our ecosystem.

Most crucially, AI is industrializing intelligence. Before this, human intelligence was never industrialized. We could mass-produce plastic toys, computer chips, but human intelligence wasn't industrialized. Now, AI allows us to amplify our own intelligence countless times.

Like electricity: when electricity was first invented, electricity itself was most precious. But when electricity could be supplied at scale, electricity itself no longer mattered, because it became standardized. What mattered was the container for electricity. A hair dryer and a computer — they produce completely different value.

Similarly, when AI makes intelligence infinitely available, content itself no longer matters. What truly matters is the container for content: the brand, media outlet, or platform that carries it. This is McLuhan's "the medium is the message" — put plainly: the container of content is greater than content itself; the container of intelligence is greater than intelligence itself.

Our generation of creators — what we truly need to do is design containers for intelligence, design containers for AI. Just as ByteDance transformed the entire content ecosystem through information distribution, I now see an opportunity comparable in scale to ByteDance: we can use AI directly to fundamentally transform the content ecosystem on the production side.

For creators, this is a choice: either stay in your marsh, or connect yourself to the Three Gorges project, transforming your work into using AI to design workflows, to design containers for intelligence.

Just as electricity moved industrial production from handicraft workshops to assembly lines, AI is moving content creation from individual studios to intelligence factories. When an ordinary person can spend a few hundred dollars a month to access virtually unlimited computing power, what matters is no longer writing one good article, but designing a container that can continuously produce 90-point content.

How to build more advanced intelligence containers, how to make AI-produced content carry more personal imprint — topic selection, emotion, and exclusive case studies — these are creators' core competitive advantages.

Finally, this very article was dictated by me over 40 minutes, with DeepClaude helping structure it and writing it in segments. Total cost: $2. I only modified a few words and added the headline.