You could just change it yourself. So why ask AI to do it?

One word in a page title needs changing.

The file is right there. Open it, Ctrl+F, change it, save. Thirty seconds.

Ask AI? I'd have to type out instructions, wait for it to run, then check whether it got it right. Any way you slice it, that's slower.

I still ask AI to do it.

And my hands stay right next to the keyboard, resisting the urge to touch it.

One word is never just one word

A real example from last week.

We lowered the price of the LINE booking system on our website, from NT$45,000 to starting at NT$20,000.

Just one number.

I told the AI one sentence: "Change the booking system to start at 20,000." It ran for a few minutes, then reported back:

49 files changed.

I scrolled down the list, line by line:

The Chinese services page. The English services page. Five service cards on the homepage. A dozen or so blog posts in Chinese and English, including derived claims like "30 times the difference" that were calculated from the price. The sitemap. The pricing data hidden in the page that only Google reads. The description file meant for AI search engines. Captions for scheduled Instagram posts. Three IG pricing graphics we'd made long ago with "45,000" printed on them. It regenerated those.

The last line was the auto-reply DM the system sends when someone comments "more info" on IG or Facebook. It said 45,000 too. That got changed as well.

I stared at the list and thought: if I'd done this myself, I would have updated the homepage and the services page, then closed the files feeling pretty satisfied.

The other forty-odd places would still say 45,000.

Until one day a client sends a screenshot on LINE: "This page says 45,000, that one says 20,000. Which is it?"

This isn't hypothetical. We've actually been caught. On a previous project, a key number changed from 142 to 983. I updated the pages I could see, missed several others, and had someone point them out to me one by one. That's a terrible feeling.

So, reason number one:

The hard part was never "changing" it. It's "finding every instance."

People only remember the places they've seen. AI goes through the entire folder.

But for me, the real reason is the second one

Even when there really is only one thing to change, I still ask AI to do it.

Because I need tobreak my habit of doing it myself。

I've been building websites and writing systems for years. My hands are fast. When I spot a problem, my hands are already moving before my brain finishes thinking.

That habit used to be my strength. Now it's my biggest obstacle.

Every time I think "this is small, it's faster if I just do it," I'm pouring another layer of concrete over the old habit. I save thirty seconds, but I never learn how to hand things off.

And handing things off takes practice:

• You have to be able to describethe outcome you want, not the steps
• You have to knowhow to check the result
• You have to be able to hold back andget it done without touching the keyboard

It's like learning to drive. You don't hit the highway on day one. You start by circling an empty parking lot.

Small tasks are that parking lot. If something breaks, the damage is minor. Once you're comfortable, handing off big things comes naturally.

22 slides: by slide 15, you're zoning out

The second example will feel familiar to anyone with an office job.

It's 4 p.m. A 50-slide deck is due tomorrow morning. Your boss walks past your desk and glances at the screen:

"That product image is too small. Make it bigger."

That image appears in the bottom-right corner of 22 slides.

You click slide one, select the image, hold Shift, drag the corner, enlarge it, align it. Next slide, same thing.

Slide 5, still going smoothly. Slide 10, your wrist starts aching. Slide 15, your eyes are glazing over. Is this one bigger than the last? Slide 22, finally done.

Then you flip through from the start to check, and find slide 8 is crooked and slide 13 was skipped entirely.

What I do now: hand the file to AI with one sentence:

"On every slide in this deck where this product image appears, scale the image to 1.5x and snap it to the bottom-right corner. Don't touch anything else."

It finds the 22 slides on its own and changes them all at once, every image the exact same size and position.

But here's something most people don't think of:How do you check the result?

If you're still flipping through slide by slide, you're back to zoning out on slide 15. So I add one more instruction:

"When you're done, give me a single thumbnail overview of every slide."

Fifty slides on one image. If any slide looks off, you spot it instantly.

Delegating doesn't mean walking away. People who delegate well also spell out how they'll check the work.

Our own website: from edit to live in one sentence

The third example is something we use every day.

We built the Satsuma website ourselves. Changing one thing used to look like this:

1. Open the file and edit the content
2. Run a preview on my own computer and check whether the mobile layout breaks
3. Commit it to GitHub as a backup (GitHub: a cloud repository for code that records every change, so if something breaks you can roll back to any previous version)
4. Push it and wait for it to auto-deploy to the server
5. Open the live URL and confirm everyone's seeing the new version

Five steps, four different tools, two or three commands to memorize. One typo at any step and the site won't load.

Even more annoying is the "done. Oh wait, need to upload again" loop. That can happen five or six times in one evening.

Now I just say one sentence:

"Change the homepage headline to such-and-such, and push it live."

Then I go get a glass of water.

By the time I'm back, it's gone from step one to step five, opened the live URL itself to check, and reports: "Live. The URL is showing the new version."

The article you're reading right now was published exactly this way.

I discussed the structure with AI, it wrote the first draft, and I told it what to change and where to push harder. Once it was final, the layout, generating the page, committing to GitHub, going live, notifying search engines, and opening the URL to confirm were all done by it.

From start to finish, I didn't touch a single file for this article.

Picture yourself as the boss sitting in the back

At this point, you might be thinking: "Sounds impressive, but I can't code."

So picture a different scene.

You're sitting in a very comfortable chair with a cup of coffee in hand. In front of you sit three people: an engineer, a designer, and a planner.

You don't need to open any software. All you do is talk.

"Tweak this headline." The engineer's done it. "Make this image bigger and put it bottom right." The designer's done it. "This page is good. Push it live." It's live. "Hmm, actually, change that back." Changed back.

An idea hits you at 2 a.m.? Just say it. Revising the same line of copy for the seventh time? Just say it. Your tone was a bit sharp? That's fine too.

You don't have to manage their feelings.

No reading the room to pick the right moment to ask, no worrying whether they'll roll their eyes at the fifth revision, no tacking on "thanks, drinks are on me" afterward.

That's the position AI can give you now. Not a search engine that answers questions a bit better, buta whole row of people sitting in front of you, on call whenever you need them。

But in this boss's seat, there are two things no one can do for you:

You have to know what you want. No matter how skilled the people in front of you are, if you can't say it clearly, they can only guess. If they guess wrong, that's not on them.

You have to sign off on the work. When they say it's done, it only counts once you've seen it with your own eyes. You're the one who signs off and takes responsibility.

So whether you can code doesn't matter.

Whether you can be the boss does.

(The "engineer who never complains" comes with its own risks. I wrote about that before:"I finally have an engineer who never complains"。)

I spent a long time as the delivery guy myself

I'm not saying this because I knew how from the start.

When I first started coding with AI, I was just a courier.

Here's the picture: the AI chat window open on my left monitor, and a remote server connection on my right (VM: a computer rented in the cloud, where the website runs).

I'd ask a question on the left, and the AI would spit out a block of code. I'd select it and copy.

Switch to the right and open nano. (nano: a plain-text editor on the server. No mouse, no colors, no undo button to mash. Every action uses keyboard shortcuts.)

Find the line to change. Paste. Ctrl+O to save, Ctrl+X to exit. Restart the service.

Broken.

Switch back to the left, copy the error message to the AI. The AI says: "Sorry, please replace line 42 with the following."

Copy again. Switch right again. nano again. Find line 42 again. But after the last paste, the line numbers have shifted. Which line is line 42 now?

Pasted in the wrong spot. Indentation broken. Missed the last bracket when copying. Accidentally pasted a line of the AI's explanation into the code.

After ten round trips, my eyes started blurring.

After twenty, I couldn't tell anymore: which version is running on the server right now? That last error, was it the AI's mistake or mine in the transfer?

By midnight, I was staring at that white-on-black screen, hands on the keyboard, with no idea what to type next.

By then, AI was already very good at writing code.

The bottleneck wasn't the AI. It was the courier in the middle.

It was me.

Later, I switched to letting the AI work directly on the server. For the same job, I was left with just two tasks: say what I want, and check whether it got it right.

That's when it clicked. The problem was never whether AI was smart enough. It was that I'd been standing between it and the work, using two hands and nano as a very narrow, very leaky pipe.

Why do most people still treat AI as a Q&A tool?

It's not that they don't know how to ask. It's thatthey haven't let AI touch their stuff.

Here's a distinction many people miss. They think these are the same thing:

Claude (chat on the web and mobile app): the consultant behind the counter

You open claude.ai and chat with it. You can upload files; it reads them, edits them, and hands you a new file to download.

But it sits behind the counter. You hand over a copy, it hands one back. The courier running back and forth in between is you.

Claude Code: the employee who sits right in your office

It works on your computer. Point it to a folder, and it can open the files inside on its own, edit, move, save, run programs, commit to GitHub, and upload to the server.

Compare using the three examples above:

In one sentence:Chat is for questions. Code is for getting things done.

If you want AI to actually do things for you, you have to let it into your workspace.

Your first task: clean up your Downloads folder

Leave your important work files alone for now.

Open your computer's Downloads folder and take a look.

Does it look like this:報價單(1).pdf、報價單(2).pdf、報價單_final.pdf、報價單_final_真的最終版.pdf. Meeting photos from three years ago. Images saved from some LINE group you can't remember. Installers for software you stopped using ages ago. Thousands of files crammed together. Every time you need a contract, you scroll for minutes, give up, and just ask the other person on LINE to send it again.

This folder is the best first task for practicing being the boss. The reasons are simple:

• It's messy enoughthat you can tell at a glance whether AI organized it properly
• It doesn't matter, so if the cleanup goes wrong, it won't hurt. Most of what's in there is stuff you forgot to delete anyway
• It's genuinely annoying, so you'll feel the difference the moment it's done

How to do it on Windows

The Claude desktop app has Claude Code built in. No white-on-black terminal, no commands to type.

1. Install the Claude desktop app

Go to claude.ai, download the Claude app for Windows, install it, and sign in. Claude Code requires a paid plan (Pro or above). We also recommend installing Git from git-scm.com first (a free tool for managing file versions that some Claude Code features use). Just keep clicking "Next."

2. Switch to the "Code" tab

Open the app. At the top there's a Code tab. Click it, then click + New to start a new task.

3. Choose a folder

Below the input box is a row of settings. For the first one, choose Local(work on your own computer), then pick a folder next to it and point it to your Downloads folder.

4. Start with "Plan" mode

In the same row there's a mode menu. For your first time, choose Plan: the AI will only look and propose a plan. It won't touch a single file.

5. Describe the outcome you want

"Organize this Downloads folder for me. Sort by type into: Documents, Images, Installers, Archives, Other. Group multiple versions of the same file together.Only move files. Don't delete anything. Show me how you plan to sort them first."

It'll go through the whole folder and come back with a plan: which folders it intends to create, which files go where, and how many sets of duplicates it found.

6. Review the plan, then let it run

If the plan looks right, switch the mode to Accept edits(it can move files and create folders without asking you every time), and tell it "Go ahead."

Then go get a glass of water.

By the time you're back, the folder you haven't dared to open in three years will be sorted.

Note the line in step five: "Only move files. Don't delete anything." That's the safety net a boss leaves themselves when delegating. Even if the sorting isn't to your liking, everything's still there. Just ask it to redo it.

This is your parking lot. Get comfortable here, then take it onto the highway.

Next time your hands reach for the keyboard

Pause.

Even for a thirty-second job. Say it instead.

It'll feel awkward at first. You'll think, "It's faster if I just do it." That's the old habit pulling you back.

Push through it, and you go from the person dragging images slide by slide to the one sitting in the back with a coffee.

Want to be that boss, but need someone to help you take the first step?

By now, some of you want to try it but are stuck on very practical questions:

Can I install this on a company computer? Which files can AI touch, and which are off-limits? How do I phrase things so it understands? How do I recover when it gets something wrong?

We deal with these every day. Our corporate AI trainingskips the vague talk about AI trends. Instead, usingyour own computers, your own files, and the tedious task you do every day, we walk you through it hands-on: from organizing your first folder to handing off an entire weekly recurring task.

When the course is over, you won't just know one more piece of software.

You'll have a whole row of people sitting in front of you.

→ Talk to us about a training session