Part 1 is here. In short: I wanted AI to make money for me, got told to use my eleven years of line-of-business experience, landed on legacy modernization — and then wondered whether AI can actually recover the things that were never written down anywhere.
Fine, let's test it
Which is how we arrived at:
let's have AI do the legacy modernization and see what happens.
Take a real legacy business application.
Hand it to the AI.
Hand over the documentation too.
And the historical context.
Then see how correctly it can modernize the thing.
🐼 "Great."
🐼 "…"
🐼 "So where exactly is this legacy business application's source code?"
Obviously.
Every business application I've worked on is internal to a company or a client.
I can't just publish the source on GitHub.
I can't publish the specs either.
Or the old emails.
Or any of the back-and-forth about how the business actually works.
But what I want to test here isn't simply "can it turn old code into new code."
Why did the code end up like this?
Who knew what at the time, and what didn't they know?
What questions were asked, what answers came back, and what did the developer end up deciding alone?
I want legacy that includes all of that.
🐼 "And where am I supposed to find a conveniently packaged experimental legacy system like that?"
…
🐼 "Or, wait."
🐼 "Do I build one from scratch?"
🤖 "That's it!!!"
To modernize legacy, first build the legacy
And so.
In order to run an experiment on legacy modernization,
in order to have a legacy application to run it on,
I started scenario coding.
How did it come to this. 🐼
The point is not to artificially produce finished old-looking code.
First, you reconstruct the situation the system was built in.
There's a developer.
There are people who own the business process.
There's a specification.
But the information isn't complete.
You can ask questions, and there's no guarantee the answer you want comes back.
The developer makes a call anyway, and writes the code.
Then time moves forward.
Requirements change.
People change.
Decisions accumulate.
That way, you build the process by which legacy becomes legacy.
And at the very end, you hand it to a modern AI.
🐼 "Here you go. Please modernize this."
That's how
AI Legacy Modernization Corpus — AI-LMC
started.
Next: 2008
The first thing we decided to build is a business application from 2008.
It isn't legacy yet.
At this point it's just new development.
The developer in charge is Yamada-san.
He doesn't yet know what his system is going to become.
I do, of course.
After all, I'm the one who decides.
🐼 "Good luck, Yamada-san."
The experiment starts next time.
References
- Digital Agency (Japan), August 2025 — based on the summary report of the Legacy System Modernization Committee: link
- .NET and .NET Core lifecycle: Microsoft Learn
- Polymarket geographic restrictions: Help Center
- The corpus itself, including the pre-registration: github.com/mori-ikuri/ai-lmc
Top comments (0)