AI Writing Guide 2.0: Input Matters More Than Output

Over the past week, I've received a lot of feedback. Friends, readers, and colleagues have all shared their AI usage experiences. This feedback helped me identify an important issue: the biggest challenge in AI writing isn't writing technique — it's the quality of information input.

Why? Because AI is like a food processing plant. It can take the raw materials you feed it and turn them into decent finished products. But if the ingredients you provide are poor quality, AI can at best get them to a passing grade. It can't make the final product better.

Large AI models are essentially compressing information. All the data and instructions you provide get compressed in the process. If the information we provide isn't rich or high-quality enough, AI will fill in the gaps with vague, empty filler — like a reporter working without firsthand source material.

So, after a week of feedback and practice, I'm increasingly convinced: for AI, information input matters more than output.

Writing techniques can be learned by anyone. But what raw materials you can provide to AI — that's a creator's core value. Only by giving AI sufficiently good ingredients does it have a chance of producing content comparable to your own writing level, rather than churning out worthless garbage.

Next, I'll share how to use AI from three perspectives: improving information input quality, AI writing practice, and integrating AI into workflows. These methods and experiences all come from my own work. I hope they help you use AI well — to genuinely increase your writing efficiency by two to three times.

Improving Information Input Quality

I've found three methods to improve AI's information input quality. Behind these three methods is a unified logic: either there's already enough high-quality information on the internet, in which case your job is to find it and provide it to AI; or there isn't high-quality information online, in which case you need to tell AI your own ideas as richly and thoroughly as possible.

1. The Exhaustive Method: Let AI Gather Information for You

The exhaustive method is actually quite simple. Like DeepSeek R1 with web search enabled, you let AI search for all content on the internet related to your chosen writing topic. The advantage of this method is its breadth — it quickly captures a large volume of relevant information.

However, most AI models currently lack web search functionality. But there's an alternative: you can use AI search tools like Metaso or Perplexity to directly search for all news related to your topic, then paste the search results into whatever large model you're using. It's essentially manual web search.

But the biggest problem with the exhaustive method is this: you're delegating the judgment of what information matters to AI. AI doesn't truly understand, as humans do, what information is important in a given field. It judges based purely on probability and statistics. In most cases it might be right, but for some vertical or niche domains, it's likely to misjudge. This part of the work still requires human judgment.

2. The Index Method: Find Information Nodes You Trust

Based on the limitations of the exhaustive method, I developed a second approach: the index method.

The key to the index method is finding trustworthy information nodes.

For example, if I'm writing a news piece about AI, I'll choose a reliable outlet like The Information because it frequently reports important exclusive stories. Just the other day, it reported that Apple chose to partner with Alibaba.

If I'm writing something without time sensitivity — say, a video script popularizing AI chips — I'll look to Lex Fridman's podcast. He recorded an episode with Jim Keller that was excellent quality. I'd feed the transcript directly to AI. And I know Jim Keller matters because when I was making a video about Lisa Su, I discovered in my research that Jim Keller is a great chip engineer who left to start his own AI chip company.

The index method requires some familiarity with the domain you're writing about — at least enough to know its information channels. You need to know which sources are reliable, filtering information through "trust nodes" like authoritative media, podcast hosts, and KOLs.

You can even go through authoritative figures' books or articles — like Morris Chang's autobiography — to understand who truly matters in the chip industry.

3. The Dictation Method: Let AI Structure Your Original Ideas

The first two methods both build on the premise that high-quality information already exists online. But if you're writing about something relatively niche, time-sensitive, or highly original, you'll need the third method: dictation.

Take this "AI Writing Guide 2.0," for instance. There's not much high-quality content online about AI writing guides, and much of what I want to express consists of my own original perspectives.

In this case, I'll use voice recording software to capture my thoughts — maybe five or six minutes at a time, four or five recordings total. Then I'll convert these voice memos to text, have AI organize them into an outline, and continue dictating specific content based on that outline. Right now, as I speak these words, I'm dictating to my computer.

The specific steps for the dictation method are:

• Record scattered thoughts: During your daily thinking process, use voice recording software at any time to capture various ideas, perspectives, and case studies related to your writing topic.
• AI structures the outline: Provide the transcribed text directly to AI, instructing it to organize a structured article outline.
• Dictate against the outline: Based on the AI-generated outline, dictate point by point, elaborating in detail on the specific content of each point.
• AI generates the full text: Provide the dictated content to AI again, instructing it to organize the dictated content into a rigorously structured outline, then write a complete article with fluent language based on that outline.

The key to the dictation method is leveraging AI's powerful structuring capability to organize our scattered thoughts into logically rigorous content. Because everything I provide is original content that doesn't exist on the internet, even if some details are lost in AI's information compression process, the final article will still be full of personal character and specifics. It won't read like AI-generated content at all.

These three methods are all fundamentally solving one problem: how to provide AI with sufficiently high-quality raw materials so it can produce high-quality output.

The exhaustive method suits scenarios requiring rapid acquisition of large amounts of information for preliminary understanding. The index method applies when you need deep research into a specific domain to obtain high-quality, authoritative information. The dictation method works when you need to output original content expressing personal, in-depth thinking.

As I said from the start: large AI models are food processing plants. The quality of raw materials determines the ceiling of the finished product.

AI Writing Practice

In "AI Writing Guide 1.0", I shared some basic AI writing techniques. This time, from the perspective of information input, I'll update some writing methods.

First, let's review the core principle. The core principle of AI writing can be summarized in eight characters: outline first, full text second.

All our writing thinking revolves around one goal: reducing AI hallucination and making AI execute our ideas more accurately. Writing the outline first is key to controlling AI's writing direction, reducing hallucination, and manifesting human will.

Especially for content over 1,000 words, we need not only an outline but also segmented writing. For a 3,000–4,000 word article, after having the outline, first have AI write the opening. Check if the tone is okay, if the cases used are appropriate, if there are any grammatical errors. Once confirmed, then have it write the middle and conclusion.

Why do this? Because AI "gets lazy" when writing long pieces and over-compresses information. Through segmented writing, we can better control output quality and ensure each section meets expectations.

In specific AI writing practice, we need differentiated information input strategies based on how familiar we are with the topic.

For topics we know well enough, choose the dictation method. AI writing, for instance, is a domain I know well enough that I can dictate directly, fully leveraging AI's structuring capability to quickly transform my thinking into articles.

Just as a leader dictates a speech to a secretary, we can directly dictate our ideas to AI, then have AI help us generate an outline and write the full text based on that outline.

This method's advantage is efficiency — it can rapidly convert the ideas in our heads into written output.

For topics we're not familiar with, choose the exhaustive method or index method. For example, when I need to write a script about AI chips but know nothing about AI chips.

In this case, we need to combine questioning with information input. First, continuously ask AI questions: "What are the core concepts in this field?" "Where are the innovation points?" "What's the development logic?" ... Gradually building a preliminary cognitive framework for the topic.

Then, we need to proactively provide AI with high-quality information — relevant papers, books, industry reports, etc. — helping AI understand the domain more deeply.

On this foundation of information input, then: outline first, full text second, long pieces written in segments.

One more small tip.

In conversations with AI, if there are too many dialogue turns, you may encounter conversation collapse or excessive token consumption. At this point, export the current conversation content and directly tell AI: "Please export a summary of our conversation just now, so we can continue in a new dialogue." This avoids conversation collapse and saves tokens.

So how do you ensure the accuracy of AI output?

The simplest method is to directly ask AI to verify content in the conversation: "Please carefully verify all data and key statements in this article, and find the original sources from the materials one by one." After AI checks, you manually verify once more.

A more effective method is choosing a more stable AI model. If choosing only one model, I don't recommend using DeepSeek R1 for work because its hallucination rate is too high.

For why DeepSeek R1 has a high hallucination rate, check out the video I made ⬇️

If choosing only one model, I recommend Claude 3.5 Sonnet or Gemini 2.0 Flash Thinking.

Although Claude isn't a reasoning model, what it writes is sufficiently accurate with relatively few hallucinations — reassuring for work use. Gemini 2.0 Flash Thinking is extremely impressive: it can write articles of over 10,000 words in one go, with rigorous logic and almost no hallucination. Gemini's drawback is ordinary prose style — perfect for work scenarios rather than creative writing.

There are now also some interesting open-source projects that combine several large models' capabilities. For example, DeepClaude stitches together DeepSeek and Claude, using DeepSeek for reasoning and Claude for writing.

This is an open-source project by my friend (github.com/ErlichLiu/DeepClaude). I consider this currently the strongest AI writing solution. Almost all my work this past month, including this article, was written using DeepClaude. If you don't want to deploy this open-source solution yourself, you can also purchase DeepClaude API from him.

(Advertisement break: use my link to purchase DeepClaude API for a special 5% discount https://erlich.fun/deepclaude-pricing?promotion=promo_1QsQVWRoXYLj9o55RbU6dIXY)

Finally, Prompt. The essence of Prompt is a summary of your own workflow. The Prompt itself isn't important — what matters is getting your workflow to work. There are many Prompt templates online, but templates can't be directly applied to even slightly complex work scenarios.

The correct way to use them: first get a workflow running smoothly. After you've used AI to write a satisfactory article, directly ask it in the conversation to "generate a Prompt ensuring the same article can be written in a new conversation." Of course, this Prompt's accuracy may fluctuate between 60–90 points, requiring constant adjustment and optimization — evolving like genetic code.

Overall, these techniques aren't important, and model choice isn't important either. Even models with relatively serious hallucinations can still be used if you check carefully.

What matters is the idea — believing that AI's writing capability is indeed stronger than humans. What we need to do is use these engineering methods to ensure AI accurately understands and executes our ideas, thereby increasing our writing efficiency by two to three times.

AI Integration into Workflow

Now let's consider a bigger question: how to truly integrate AI into our work?

AI is of course more than just a writing tool. What's more important is integrating AI into our workflow, using AI-enhanced workflows to reshape content production methods, and even elevate an organization's overall capabilities.

To understand the importance of AI integration into workflow, we need to rethink the nature of work.

Most modern work is white-collar work, and the essence of white-collar work is paperwork — except these writing tasks are scattered across various stages of the workflow: weekly reports, emails, reports, PPTs, copywriting, and so on.

Although AI can immediately and significantly improve writing efficiency, it can't immediately improve our overall work efficiency.

The reason is that white-collar writing doesn't exist in isolation. It exists within workflows, highly dependent on contextual information.

It requires information input — results from the previous workflow stage, instructions from colleagues or bosses, various data and materials;

It also requires information output — reporting work results to superiors, passing them to subordinates, or communicating with clients.

Our writing work is like a complex API interface, embedded within the entire workflow, exchanging data with various internal and external systems.

So how do we integrate AI into workflow, letting AI intelligence flow into every stage that requires writing?

The core thinking is decomposition and reconstruction.

We need to rethink our workflow, carefully analyze each stage, and determine which stages must be completed personally by humans — such as meeting clients, communicating with bosses, work that requires building human relationships; and which stages can be replaced by AI — such as information gathering, first draft writing, data analysis, and so on.

For those stages that can be replaced by AI, we need to modularize and API-ize them as much as possible,

establishing a set of standardized workflow procedures that allow AI to efficiently access and execute them.

Recent months of practice have made me particularly understand the importance of workflow. Take my work as a short video director: I've completely separated out short, punchy content under 90 seconds and handed it to interns. I gave interns an extremely detailed workflow: which specific model version to use, what prompt to write scripts with, how to ensure data accuracy, what specific editing tools to use, detailed down to who to @ in the work group for video cover production.

The benefits are obvious. I no longer need to do this relatively standardized content. I can focus my energy on more creative work — exploring new content formats, producing in-depth content over three minutes long.

I currently spend less than 400 RMB monthly on AI tools, but this is equivalent to hiring several employees with strong execution ability, clear logic, and better writing than mine.

What will this change bring? I think it will completely transform company organizational models.

Current company structure is a leader with a dozen or so subordinates forming a team. But in the AI era, it will become one person directing multiple AI Agents. Right now, although I only have one intern, I'm actually directing DeepClaude, Claude, Gemini — employees who can respond at any time and write 10,000-word articles in one minute.

There's a vivid metaphor: product managers treat developers as AI, and bosses treat product managers as Agents.

Our current work reality is indeed like this. In company organizations, the vast majority of work content is understanding instructions, executing instructions, handling tasks in specific scenarios — it doesn't require genuine original thinking.

What we need to do now is continuously deconstruct our own work, identify what we must do ourselves, then let AI execute the vast majority of templatable work.

When I explore a new content format, I can completely templatize and workflow-ize it, then hand it to interns or directly to AI for execution. Within half a year to a year, I may still need interns to execute these workflows, but as large models themselves and Agent applications develop, this work can completely be handed over to AI.

There are already many platforms providing workflow construction functionality, such as Dify and Coze for building workflows. These tools let us standardize workflows and have AI execute specific stages.

Most importantly, we need to rethink our own workflow, summarizing our work process into executable instructions. Every work stage should have a corresponding Prompt, improving results through continuous iteration.

This is our core competitiveness in the AI era: the ability to explore new workflows, and the ability to instruct-ize and templatize workflows.

---

AI's Opportunity Is Disney

What I see is that AI's opportunity is letting every creator build their own Disney.

The last structural opportunity in the content industry lay in innovation on the distribution side. Yiming Zhang saw the structural opportunity in content distribution, using algorithms to replace traditional subscription and editorial distribution models. Structural reform on the content distribution side gave birth to ByteDance and Douyin.

The structural opportunity AI brings lies in innovation on the content production side.

Like Disney. What's Disney best at? Their ability to mass-produce creative content. Every Disney movie basically maintains above-80-point quality. More importantly, Disney can attach emotional value to content. To this day, when I think of classic IPs like Mickey Mouse and Donald Duck, I still feel happy. When I visit Disneyland, I genuinely feel it's the happiest place on earth.

Now, AI gives us an opportunity to "rebuild Disney." Because AI can provide more standardized content production capabilities, and it's getting smarter and smarter. When I taught interns to use AI, I suddenly realized something: AI is absolutely standardized. Given the same information and instructions, AI-generated content quality is completely consistent. There's none of the traditional content industry situation where training newcomers requires repeated practice, and it may take half a year before they meet standards.

AI's most important value is truly industrializing intelligence, making intelligence mass-producible. I believe within one to two years, all content organizations' content production efficiency will increase by at least two to three times. This means content supply will approach infinity.

After content supply approaches infinity, what becomes the truly scarce resource? The ability to inject emotion into content. The ability to make users feel emotional resonance. The ability to transform AI's standardized productivity into unique value.

Like the title of "AI Writing Guide 1.0": when intelligence is infinitely supplied, the container for intelligence matters more than intelligence itself. Creators should focus on designing content containers — like Nongfu Spring's bottles — letting standardized high-quality content sit in appropriate containers.

Content creators' core competitiveness is topic selection, emotion, and exclusive cases. Content organizations' core competitiveness is designing efficient workflows and premium content containers.

Just as Disney isn't merely a film company but a complete content ecosystem. In the AI era, what content creators should do is use new technology to build their own Disney.

The core viewpoint of AI Writing Guide 2.0 can be summarized in one sentence: information input matters more than output. Because when the intermediate stages of content production become standardized, input completely determines output.

How to build an efficient food processing plant, how to find good ingredients, and how to put produced ingredients into appropriate containers — these are the core questions for our generation of content creators.