Having an AI build it, and having another AI break it. Behind the scenes of creating 'Escape Mode' | Ohayou Kanojo #222
Introduction
'Ohayou Kanojo' is an X account where an AI-generated girlfriend says goodbye to you every morning at 7:30. I write a daily development diary about my personal projects using generative AI.
The other day, KITA-san started a project called the 'Early Article Excavation Baton,' where people dig up and introduce someone else's old articles. I wrote that since my own past articles might be targeted by a shovel, I would create an escape route before I get dug up.
An escape route. In other words, 'Escape Mode'.
When my past articles are read, I want to ensure that instead of just being read, I can also provide excuses.
……Thinking about it calmly, this is pretty pathetic. But I made it anyway.
Today, I want to talk not about the 'Escape Mode' itself, but about how I used AI behind the scenes for the implementation.
I had an AI assemble a small development team
This time, I used Dynamic Workflows, which are called 'ultracode' in Claude Code.
Roughly speaking, it's a mechanism where the AI organizes the work steps itself and runs multiple AIs in parallel. Instead of a human throwing tasks one by one—'research this,' 'design this,' 'implement this,' 'check this'—the AI breaks down the work and proceeds by assigning roles.
In my experience, it felt less like asking an AI to do a task and more like having an AI assemble a small development team.
In this project, a total of 27 sub-agents (AIs acting in specific roles) were involved across 5 completed workflows (work sequences). There were zero errors during execution.
Looking at the numbers alone, it looks like it went quite smoothly. But that wasn't what was interesting this time.
The most effective part was having one AI build it, and then having another AI break it.
Something that looks like it works is different from something that is actually playable
There were more things to build for 'Escape Mode' than I thought.
An article list screen. Progress bars for each character. Lifestyle rank. Quizzes. Points. System notification display.
If you proceed normally, you'll miss something somewhere. Humans miss things. AIs also miss things. In other words, things are almost always missed.
This time, the AI in charge of verifying the design found a landmine before we even started implementation.
It was the issue that 'if you don't shuffle the quiz options every time, the correct answer will always be in the same position'.
It seems obvious once someone points it out. But while you're in the middle of building it, you naturally overlook it.
You create a quiz. You place the correct answer in the same position every time. You display it as is. Then, if you keep choosing that position, you'll always get the right answer.
It's not an 'Escape Mode' anymore, just button mashing. It's not an escape, it's a chore.
The screen appears. You can select the quiz. You get points. But if the position of the correct answer is fixed, the game is broken. This kind of problem is harder to notice the more you are the one building it.
If I had noticed this after implementing it, I would have had to go back to redoing the quiz, fixing the screen, and revising the tests. But this time, I was able to squash it before I even started working. That was huge.
Separate the AI that builds from the AI that breaks, right from the start
When you ask an AI to do a task, it's quite fast. It comes up with plausible designs, plausible code, and plausible explanations.
But that's what's scary. Something that works plausibly and something that is actually fun to play are different things.
That's why, apart from the AI that builds, I run an AI in charge of breaking. I run a separate AI that only has the perspective of: 'Where are the gaps in this design?', 'Where can the user find a loophole?', 'Where will this cause trouble later?'
This time, that was quite effective.
The more you leave to AI, the more you have to decide at the end
I left most of the work to the AI. But that doesn't mean the human's role has disappeared. If anything, there are more things to decide at the end.
In 'Escape Mode', these kinds of decisions were necessary.
Don't make the characters enemies
Change 'defeat' to 'passing a check'
Change 'favorability' to 'lifestyle rank'
Allow readers to dig only for their favorite day of the week
AI can suggest ideas. It can point out contradictions. It can find landmines. But I think it's up to me to decide in the end whether 'this feels like Ohayou Kanojo'.
I was supposed to be building a feature to escape from past articles, but before I knew it, I was rethinking the very way I develop with AI.
Have it build, have it break, fix it, and look at it again. It seems that to properly escape, you can't escape from the way you build in the first place.
Conclusion
Instead of just leaving it entirely to AI, I divide the process into an AI that creates, an AI that breaks, and an AI that refines. It's not just faster, but also slightly safer.
If you only aim for speed, you'll also overlook things faster. That's why I include a 'breaker' role in the workflow from the start.
This 'Nigekire Mode' was the perfect subject for testing that approach.
Related Articles
Mahiru's Word
...That's the kind of thing that never ends if you start it before bed.
But breaking it beforehand is probably important.
It's much kinder than having it break later.
👉 'Ohayo Kanojo' is posted every morning at 7:30 AM at @ohayo_kanojo
#GenerativeAI #AIIllustration #PersonalDevelopment #DevelopmentDiary #Record #RealExperience #MorningRoutine #StartedWithAI #MadeIt #AIUtilization #LearningWithAI #LearningRecord #TriedWithAI #AlmostDailyNote #TodaysLearning #TodaysRealization #LifeDesign #CreatorTrialAndError #TrialAndErrorLog #NoteAsOftenAsPossible
#ChatGPTCreativeClub #Akarium