Two weeks of talking AI into doing my work

I barely type anymore. I talk, and an AI agent (Claude Code, pointed at a folder of my notes) goes off and does the work. Here's what it actually did for me over two weeks, scene by scene, plus the places where it still needs a human watching.

A few days ago I went to Google DevFest Shanghai. I recorded three talks and took twenty-odd photos of slides. When I got home I said one sentence to my computer, and the AI pulled the transcripts of all three recordings out of DingTalk (Alibaba's workplace app), copied the slides from the photos page by page into text, matched them against the recordings, saved everything into my knowledge base, and then went through what I'd learned that day with me.

I used to go to tech conferences, take photos, and leave them in my camera roll, never to be opened again.

Over these two weeks I've used AI like this more and more: no typing, I just speak a paragraph through my phone keyboard's voice input, it goes off and works, and at most I press send. A few years ago I asked Siri or Xiao Ai (Xiaomi's voice assistant) about the weather and the time. Now one sentence gets a whole chain of things done. My biggest takeaway: the more I've stored up in my knowledge base, and the lazier I get about handing things over, the more it can do for me, and the smoother it gets. Like a flywheel.

Here's what it did for me these two weeks, scene by scene, and where it still needs me watching.

Working directly inside Feishu and DingTalk

When I used AI to write documents before, it wrote in the chat box, I copied it into Feishu (known as Lark outside China; think Notion plus Slack), then fixed the formatting.

Now Feishu has a Feishu CLI (command-line interface), and DingTalk has one too. Install it on your computer, authorise it once with your own account, and the AI can operate them directly. I say "write me a ...", and it creates the document in Feishu with the title, body and tables all laid out. Edits happen in the conversation too: it handles the big chunks, I fix the small details.

What surprised me most was the diagrams. I talked through my thoughts on "how people and AI divide the work inside a piece of software", and it turned them into a Feishu document with three diagrams: how one task moves through the software, which step should stop and let a person look, and how a beginner and an expert can share the same flow. The logic is clear at a glance. (The screenshot below is an English copy I had it make for this post.)

Drawing diagrams has always been my weak spot. I used to admire how good my former manager was at it. Now, with AI, I've more or less "stolen" the skill: my hands still can't draw, but my mouth can. When a diagram is off, I point at it and tell it what to adjust, and after a few rounds it looks right.

On the DingTalk side, I bought DingTalk's AI recording device. I used to download the recordings myself, click export, then tidy them up. Now I can't even be bothered to click download. The three DevFest recordings at the start came in exactly this way.

Writing: the ideas are mine, it does the writing

What I dread most is putting things into words: writing project READMEs, writing articles, filling in forms.

For a README, I tell it the points to cover: what the core idea of the project is, and why I wanted to build it in the first place. I know those well; it drafts from them and I review. What comes out is often better organised than what I'd write myself.

Forms are the same. Last year when I applied to AdventureX (a hackathon), filling in the application form nearly made me give up: I'd filled in this information many times before, so why write it all again? Now my experience lives in my knowledge base, and it picks the right parts for each event and writes from them. Same with résumés: one set of experience, written into different versions for different companies and roles.

Articles too. This one was written that way: I talk, it writes, I edit.

After writing: new versions, then sending them out

Finishing an article is only the start. The same article lands on different platforms with different readers, so it has to change shape: the blog gets the full piece in Chinese and English, also read by search engines and AI; the WeChat Official Account (WeChat's blogging platform) is for friends in China and needs new formatting and a cover; Xiaohongshu (RedNote) is for people scrolling on their phones, so it gets split into a set of image cards; LinkedIn is for peers and recruiters, where a short first-person post is enough; Dev.to needs an English version that points back to the original on my blog. That's the version you're reading.

Then it all has to go out: submit to Baidu for indexing, push into the WeChat draft box, fill in the LinkedIn post box, publish to Dev.to through its API.

No single step is hard, but together they're exhausting, so exhausting that sometimes I just don't post. Now I hand over both parts: it makes a version for each platform, I take a look; it publishes where it can, and for the rest I just click publish. I only write.

Our lives keep getting more complicated, with more software to use and more steps to follow. AI can pull some of that complexity back for us.

Helping me digest information

Besides conferences, it also does these for me:

• Scheduled searches: every day and every week it finds interesting people for me and keeps an eye on the latest trends and hot topics in tech, what people are talking about.
• Filtering: podcasts, blogs, news. It goes through them first and picks out what's worth my time. Long podcast episodes get boiled down into a few sections of key points, and I no longer have to read very long articles myself.
• Reading: in August, the Shanghai Book Fair had a few events on reading with AI. When I got home I had AI read the unfinished how-to books on my shelf first, and wrote on my phone at the time: "ai read the unfinished howto books that were taking up my headspace, now I can read literature". Now when I read, I just say whatever comes to mind and it notes it down; when I finish a book, we go through the notes one by one. While I was reading Paris: A Literary Companion, it turned the places in the book into a map of Paris:

Then there's Tiaodao. Tiaodao FM is a literary podcast, and its curator Cai Xin and I once thought together about how to turn it into an app or a website, so the episodes could live on in another form. The first step was the transcripts: nearly three hundred episodes all turned into text, then AI grouped them by theme, cleaned up author and book names, and worked out how the people and books connect. There's no way I could have handled that much on my own.

Building things from scratch

• Hackathon: Dev.to ran a challenge from Presage. From seeing the challenge, to having it explain how the SDK works, writing the demo, recording the video, writing the article, pushing to GitHub and submitting, it took three hours in total, and I ended up with Pip, a talking octopus that lives on my desktop. Before, just pulling the submission materials together took several hours.
• Video: I've made explainer videos with NotebookLM, and these two weeks I tried HyperFrames. It edits video with code: video editors want a drag-and-drop interface, but AI is better at writing code, so tools like this suit it better.
• Website: switching my blog's framework, adding analytics, making it bilingual, all done while chatting.

The most work went into my portfolio. The projects I've built and events I've taken part in used to be scattered across hard drives, old notes and various platforms. First I had AI organise them into a list in my knowledge base, then generate the blog's Projects and Events pages from that list: 29 projects and 17 events now, each in Chinese and English, with cover images it drew in code. Every project and every event has a detailed write-up: what I did, why, and the trade-offs along the way, plus wireframes, slides and event posters dug out of my hard drives. There's a lot in there.

Where it still needs me watching

You can't just hand things over and forget them, especially with content.

• Video: with the HyperFrames videos, I let it start straight away and didn't step in, and in the end I still had to rewrite what each lesson said myself. The free voice-over wasn't great either.
• Audio: in July I built a meditation app and had AI generate the guided meditation audio. It got made, but it wasn't good: I still had to go through the tone and wording sentence by sentence myself.
• Review: a lot of the time I'm still spending time watching it work and checking what it hands back.

With content, it still has flaws. In the end it comes down to taste: everyone's taste is different, and it doesn't know mine yet. This step needs a person watching.

What I've learned from using it

First, the more I've stored up, the better it works. I tell it about my experience once, and résumés, application forms and my blog portfolio all come from there. Change one place, and it updates the others to match.

Storing things up isn't much work either. While we're chatting, I often remember something I'd never mentioned before. I say it in passing, and it writes it down and adds it to my experience. While writing this article I only just remembered the meditation app I built in July, and as soon as I said so, it was recorded. Things from long, long ago get dug up bit by bit this way too, which feels a bit magical. All I have to do is talk.

Second, teach it once and it remembers. Work I already know how to do, I just hand over. If it gets something wrong, I explain how it should be done once and it fixes it; the fixed approach gets saved as a SKILL, and next time it follows that. WeChat formatting, Xiaohongshu cards, and digesting podcasts and talks all have SKILLs now. How my knowledge base is set up is in My personal agent is just a folder.

Third, throw it over first. Just say "I want to do this", and let it work out how.

Technology seems to move forward because people want to be lazy. As for what people will do once so much has been handed over? That's a good question...