How I Teach Ordinary People Around Me to Use AI

No.
281
Date
Read
13 min
Words
2,766

What can happen in seven months? For a student, it's one semester plus a summer break. For an office worker, it's just another Spring Festival coming around.

How it feels depends on where you are. If I worked on an assembly line in a factory, seven months might feel no different from one day.

But if those seven months are full of things happening, lessons learned, and things witnessed, it feels like a lifetime ago. That's how the past seven months felt for me, deep in AI.

Seven months ago I was still having the simplest conversations with Gemini in a browser, telling it about my problems. Seven months later I've been working with all kinds of agents1 for a long time. They were all rough around the edges at first. Now Codex has the most powerful computer use2 in the universe, and Claude Code has become the most capable and most tasteful coding agent.

The models changed just as fast: from GPT-5.4 to GPT-6.2 Astra, from Opus 4.6 to Fable 5.2, from Gemini 3.1 Pro to Gemini 4 Carbon.

Seven months ago, when I wanted a personal blog, Gemini had to explain Hugo, GitHub, API tokens, and Cloudflare3 to me piece by piece.

Seven months later I understand most of those concepts, but I never need to understand them again. I only need to say: I want to build a personal blog.

After that, there's almost nothing left for me to do.

In those seven months I used AI to start my first paid AI novel community and solved the problem of making money from fiction. At work I optimized processes and documents, and wrote a logistics management system for my department that the whole department now uses. In life, on a trip to Shenzhen a couple of months ago, Diandian, Xiaohongshu's AI assistant, was my right hand.

AI knowledge, practice, and news pile up in my head, so every time I meet someone interested in AI, it feels like finding treasure.

Liu Zhenyun said4 that a person's whole life, stumbling along, is really a search for someone to talk to.

No need to make it about the meaning of life. Even in daily life, finding someone to talk to is already hard.

Looking back, seven months ago, telling Gemini about my problems sentence by sentence in that chat box, I was in a sense looking for a "person" I could talk to.

I work in after-sales in the IVD5 industry. Most days I deal with circuit boards and people, not many documents. Among jobs hit by AI, mine is one of the least affected.

So most of my peers don't understand AI. AI can't tighten a screw for us, and it can't fix a circuit board.

But I still try to bring work and AI together. Client management, document management, and production are all handed to AI now, and it has freed up a lot of my spare time.

At work, though, it's hard to find someone to talk to about AI.

But slowly, people started coming to talk to me.

The first was Lex, a cycling friend. I've joined two rides he organized in Chengdu. He's a photographer and always gets great shots of us.

Offline I barely talked to him. I was too busy striking poses without mercy.

Online, in the group chat, when I occasionally shared something about AI, he would always join in. That's how I learned he had just started with AI. He used Codex for some things, but had no complete workflow, wasn't familiar with models or technique, and was still groping around in the fog.

Later Claude Opus 5.5 came out, and the internet was full of videos this model had made straight from code6. It was everywhere for a while. My friend Yuehui opened an account posting videos made with Claude and gained tens of thousands of followers in a short time.

Lex works in marketing at a medical aesthetics company and studies video anyway, so he came to ask me how to use Claude. I just told him what I'd figured out: what kind of account risk-control problems overseas products run into in China, and how to deal with them.

Soon he was on Claude, lost in its power. But at that point he had only switched to a new model. What was the essential difference from editing videos with Codex before?

None.

Still, things seemed to be getting a little better. Recently he told me he wants to break down the walls between departments and let everyone create on Claude's soil.

His idea: people in each department send requests to Claude from their own computers over the local network, Claude receives all of them, then hands them out to different models to do the work.

I admire his imagination. I'd seen that architecture long before, but I learned it from someone else. That he came up with the shape of it himself is rare.

He just didn't know how to make it real. I told him it's really about using one strong model as the brain, the decision layer, which takes every request and dispatches it to different models in the execution layer7.

It's a good idea, and if it lands it could change how they work. But the picture is plump and reality is bony. To pull it off he'd have to research every department, work out what tasks exist and whether they have order and dependencies, before he could really start designing.

The whole time, I didn't do much for Lex. I just answered his questions. And almost all of those questions were ones I'd asked other people or worked out myself over the past seven months.

Back then each one was as hard as climbing to heaven. Now, explaining them to Lex, I barely break a sweat.

In Lex I see myself seven months ago: an ordinary person who meets a brand-new technology and wants to understand it as much as possible, so he keeps tinkering, keeps groping.

Seven months ago I was the one looking for someone to talk to. Seven months later, other people come looking for me.

But people like Lex, willing to tinker, are still rare. Overseas products like Claude and Codex are something only a small slice of people in China can get to on their own. Most people use Tencent's WorkBuddy, Doubao, and Doubao Work, which launched recently.

Just today, WorkBuddy announced two numbers: over ten million users, and more than 100 trillion tokens8 called per day.

That means AI is long past being a toy for a few geeks. Tens of millions of ordinary people have put it on their computers. But between using it and using it well there is still a long road.

I've met plenty of people who know something about AI but haven't really put it into their work. Their worries are mostly the same.

First, it costs money. They don't want to spend a cent on AI. Of course, maybe they just don't want to spend a cent on work.

Second, they don't know how. Most people are still chatting with Doubao, maybe drawing a picture or making a slide deck at most, using DeepSeek to tidy up some material. That's it.

And many people, before they even start tinkering, hear all the talk online: don't depend too much on AI, protect your attention.

I want to say two things.

First, ordinary people have a huge amount of work that should be handed to AI.

Second, ordinary people don't have that much that needs handing to AI.

This talk of protecting attention has little to do with most ordinary people. Most people's work is very concrete, and a large part of it is meaningless.

That's the "huge amount of work" in the first sentence. As long as you're inside an organization, at the organizational level or the individual level, there is a mass of junk work: it creates no value, yet it has to exist.

For example, repetitive paperwork: weekly reports, monthly reports, quarterly summaries, application materials, briefing materials, inspection materials, and meeting minutes. (Many people have probably already felt how handy Tencent Yuanbao's AI meeting summary is.)

For example, coordination and managing up: telling the same progress update to your boss, your boss's boss, and the partner department.

For example, slide decks: three sentences of content laid out over twenty pages, with images, alignment, and it has to look good.

And performative work: looking busy, replying instantly, sending an email or writing a report before you leave; Moments-style internal PR pieces, event summaries, "experience sharing."

There's always one or two words there that bring back a memory. And all of it can go to AI.

A bit about myself.

I've been using Grok BOT a lot lately. One bot is called Meeting Minutes Officer. After every Tencent Meeting I send it the minutes from Yuanbao AI, and it records the date and the details. If I want to know what was said in a meeting, I don't dig through the recording. I just ask: what did we talk about last time?

Of course, most of what's said in meetings is junk, and you never use it afterward.

But with this bot I don't have to think about any of it, and it saves me a lot of attention and energy. That is the simplest dividend AI gives ordinary people.

Another example: calibration reports.

I often issue calibration reports9 to hospital clients, over a hundred a year. Before, there was no system, and it was exhausting. But apart from the data, these reports are mostly the same.

With AI, I standardized the whole thing.

One, the template. The company's old templates were uneven, came in several versions, and weren't good-looking. I had AI design a professional, clean standard layout.

Two, the process. Before, I prepared the data, found an old document and saved a copy, then filled it in line by line, staring at a hundred entries for half a day. An hour was normal. Now I send all the data to AI, it extracts and fills it in automatically, and produces a report that is professional in both data and appearance.

Three, management. By now there are no fewer than 1,000 reports from the past few years. I used to manage them by year myself, not fine enough, not steady enough. Now AI does it, sorted by year, hospital, report type, and instrument model, and new files archive themselves.

How convenient is that? By rough count, AI saved me at least 100 hours this year!

Is it hard? Not at all. You don't need overseas models or agents. WorkBuddy and Doubao Work can do it.

The second person who came to talk to me was a colleague in the company's marketing department. Unlike Lex, she doesn't tinker and doesn't want to. She is the most typical kind of ordinary person.

A while back she came to Chengdu on a business trip and we met. The first thing she did was ask me to "AI" a slide deck for her, because I seem to be good at AI.

So I talked with her.

What deck?

A client training deck.

How are you making it now?

Doubao.

What don't you like about it?

The content and layout are okay, but some details feel a bit off. For example, this section, it made seven subheadings. I think that's too many, too complicated.

How did you ask Doubao to do it?

I gave it a document.

Where did the document come from?

DeepSeek wrote it.

How did DeepSeek write it?

From an image.

Where did the image come from?

Doubao made it.

…How did Doubao make it?

I wrote a paragraph myself and sent it over.

Only after all that did I see it: from the paragraph she first wrote, the material had passed through Doubao, then DeepSeek, then back to Doubao.

A simple, clear request had turned into a pile of clichés between two AIs, and the deck came out mediocre.

Does she use AI? She does, and she'll even chain three tools together. AI can definitely help her. She just never thought about how to get AI to help her better.

She happened to have her laptop, so I spent the afternoon helping her install WorkBuddy.

While it was still downloading she asked, "This isn't going to cost money, is it?"

I said, "Not for now. Your account has 2,000 points, and Hunyuan 4 Flash10 is still free."

I showed her briefly how to set up a workspace and use different folders for different projects. The important part, of course, was teaching her to make a deck with AI.

Making a Slide Deck With AI in Four Steps

Step one: gather the material. Think like a person: if you were making this deck yourself, what would you need? Images, documents, put them all in one material folder and let AI read it. Give it the source material all at once, and stop passing it between several AIs.

Step two: talk it through. Don't rush AI into working. Talk through the request in depth: what should the deck look like? What content rhythm? What design style? Keep talking until you feel it's roughly there.

Step three: get a demo. Have AI make a demo first, and see whether the overall direction is right.

Step four: revise point by point. Look at the demo, find what you want changed, and change it bit by bit. For example, if she thinks seven subheadings are too many, she just tells AI to merge them into three or four.

In the end you don't just get a deck. You get your own process for making decks with AI. The general fixes and taste settle in, and next time it's twice the result for half the effort.

The whole time she asked me several more times: this isn't going to cost money, is it?

I didn't know whether to laugh or cry. I kept explaining that it's free for now, and kept talking.

But I also told her that WorkBuddy's basic membership is 70 yuan a month, and it's really worth trying. AI really can do a lot.

When I first got into AI I resisted paying too. Later I paid so I could use the top models. Paying didn't make me a hostage to AI. It opened a door to a new world.

I sincerely suggest everyone who works in an office try a product like WorkBuddy, and find every way to hand it most of the work on their computer. Keep at it, and you'll see clearly which things you really don't need to worry about anymore, and which things you still have to do yourself.

And of the work that settles out and truly has to be done by hand, there isn't much left. But that may be the part of the work that is actually valuable.

Seven months ago I was looking for a "person" to talk to in a chat box. Seven months later, AI has said the things I didn't need to say, and written the materials I didn't need to write.

The time left over is just enough for me to do that small thing that really matters, and just enough for me to sit down and have a proper conversation with someone.


Footnotes

  1. Agent: an AI program that can break a task down, call tools, and work through it step by step, not just chat with you. ↩

  2. Computer use: letting AI see the screen, click, and type like a person, operating the computer directly. ↩

  3. Hugo is a static blog generator; GitHub is a code hosting platform; an API token is a key that verifies your identity when calling a service; Cloudflare provides website hosting, acceleration, and other services. ↩

  4. Paraphrased from Liu Zhenyun's novel One Sentence Is Worth Ten Thousand. ↩

  5. IVD: in vitro diagnostics. A medical field that tests blood, body fluids, and other samples to help diagnose disease. ↩

  6. Code-generated video: the model writes code to produce the video directly, without editing software. ↩

  7. This "brain assigns tasks, several models execute them" architecture is usually called a multi-agent system. ↩

  8. Token: the basic unit of text AI processes. You can roughly think of it as how much text AI reads and writes. More calls mean AI is being used more. ↩

  9. Calibration report: a report issued after an instrument is calibrated, proving that its measurements are accurate and reliable. Hospital testing equipment usually needs regular calibration. ↩

  10. Hunyuan: Tencent's in-house series of large models. ↩

Licensed under CC BY-NC 4.0. Please credit the author and link to xiluluke.com/en/281

Xiluluke

A dialectical minimalist writing about life, reading and thinking.

Comments

0 / 2000Plain text · Line breaks preserved
Comment as

Loading comments…

    Your browser remembers your name and email once you post.

    Related writing

    WeChat · Xiluluke

    WeChat QR code for Xiluluke

    Scan with WeChat, or long-press to save the image.

    Open original ↗

    Buy me a coffee · Xiluluke

    WeChat reward code for Xiluluke
    WeChat
    Alipay code for Xiluluke
    Alipay

    Scan with WeChat or Alipay, or long-press to save the image.