I asked an AI to 'collect 20 popular articles' and it failed. The problem wasn't the AI, it was my order form

Behind the scenes of a pizza shop, Episode 2. You can read this even if you haven't read the previous one.

“AI agents are supposed to be amazing”—I get it. But I don't know what to ask them to do. Until last week, I was half that way myself.

I'm in my 40s and run a pizza shop in Gamagori. Last time, I wrote about how I'm rebuilding the inside of my PC like a company (it's okay if you haven't read it. In short, it's about building a 'factory that leaves tedious back-office work to AI').

First, just why am I building a factory? Because a shop where the owner's time is dissolved in paperwork and record-keeping has no future. What will I do with the free hands left after leaving the back-office work to machines—I'll write about that at the end.

You don't know what you've built until you run it. So, on Sunday, I did a test run.

Test run topic: 'Collect 20 popular articles'

I chose the topic of studying how to broadcast information. I decided to have an agent collect 20 popular articles on Note about the currently trending AI 'Claude Fable 5'—a new model released just last month, currently in the middle of a 'festival period' where it's effectively unlimited use on a flat-rate plan until July 7th.

All I did was ask the chat in one sentence. A few minutes later, I had a list in a file with the number of likes, comments, and links. This is the first step of the factory, 'Investigation'.

Step 2: Just mutter with your voice while reading

Next, I opened and read the 20 articles one by one. This is where the self-made transcription app I introduced last time comes in.

I just speak out loud what I think while reading. 'This title is good,' 'These are all old articles, I failed in how I asked.' When I stop the recording, the mutterings become 'impression memos' with timestamps and are automatically placed in a fixed folder.

And the third step. I ask the agent, 'Make this memo into a draft for the next article.'—The article you are reading right now is the first product of that process.

By the way, even without a self-made app, you can experience 80% of this factory starting today. (1) Record your impressions of what you read directly into your smartphone's voice memo. (2) Copy the transcription (it comes out with standard features on iPhone). (3) Paste it into an AI chat and ask, 'Make this into a draft for an article.' That's it. What I'm doing is just a version where the folder organization is automated.

What I gained from the test run: How to read numbers

Since I went to the trouble, I'll leave behind what I learned from reading 20 articles.

First, the number of likes isn't just the power of the article. If someone with 90,000 followers writes, even a quick memo gets big numbers, and there are articles with double-digit likes even if the content is good. Read likes as a set with the number of followers. I think this will be a safeguard for people who are going to start broadcasting so they don't get swayed by numbers.

For titles, lean more toward what the reader wants rather than what the writer wants to say. People are hooked by words of the 'work that takes 2 months finished in 1 day' type.

Also, timing. Someone published 'How to master Fable by the July 7th deadline' on a Friday, and I thought that was clever. If you put out the same content on a Sunday, it's already too late.

One last thing. To be honest, articles published exactly as the AI wrote them are transparent. You can't outsource your own impressions. That's why I include my thoughts as a 'voice.' That is the one part that must not be mechanized.

Test run bug report: Resolution of the order

What's important here is that a bug was found. However, it wasn't the machine that was broken, but my order.

When I asked for '20 items in order of popularity,' a lot of articles from a month ago were mixed in. That's only natural. Since older articles have accumulated more 'likes,' a popularity ranking favors the past. While reading it, I muttered, 'Next time, I'll narrow it down to more recent information. I messed up a bit.'

Next time, I'll order it like this: 'Narrow it down to the last two weeks, and include both the number of likes and the number of followers.'

An agent works exactly as ordered. But it won't pick up on things you haven't ordered. Therefore, the resolution of the order becomes the quality of the product. Once you've built the factory, the next job is learning how to write the order forms. This was the biggest takeaway from the test run.

Actually, this was the same thing I've been saying in my main job of promoting DX. What you leave to AI is the task. What you must not leave to it is the 'purpose.' I was doing in my own factory what I've seen so many times at companies: mass-producing 'plausible-looking output.' I had a little laugh at that.

The star is, after all, the pizza

While this factory is running, my hands are free. I use my free hands for dough research, developing new menus using ingredients from Gamagori, and spending time with my family. The reconstruction work is for that purpose.

To be honest, what I want to create in this shop is not just a place where the pizza is delicious. It's a cycle where people say, 'I want to buy from you'—leaving everything that can be done by machines to machines, and keeping only the work that people choose for a reason, like 'I want to go because [Name] is there.' I believe that is the kind of work that will increase in value in the age of AI. I will write about this properly someday.

The third episode will probably be after July 7th. I'll report back on how the factory is doing after the festival period is over. (I forgot to post it as a human, so it's delayed lol)