vivivi.games

← Dev log　|　日本語版

The parts that were genuinely hard, and what I did about them

Article five. The tidy version is in the first four; this one includes the complaints. Working with AI brings kinds of difficulty I hadn't anticipated.

The app was too heavy to use

The game is a single HTML file of about 1 MB. Try to handle that in an AI chat app and the app freezes. Attaching the file stalls it; long conversations make it worse. In the early days the fight wasn't "is the AI smart enough" but "does the app respond at all".

What I did: stopped putting files in the chat. Files live in GitHub and the AI works there. The chat is for decisions and instructions. The GitHub integration in article one was, among other things, a performance fix.

The AI ran ahead on guesses, repeatedly

This was the most frequent problem. Some specifics:

The common thread: an AI would rather fill in something plausible than say "I don't know". There's no malice in it. It's just good enough at filling in that you don't notice.

What I did: enforce the "unverified" label. Ask whether it checked. If not, make it check before continuing. Tedious — but cheap compared to the number of times I got burned.

I moved to ChatGPT, and that was its own hell

Worn down by the weight and the guessing, I switched to ChatGPT for a while. The honest conclusion: a different kind of hell.

My candid impression is that quality dropped noticeably compared to Claude. This field moves fast and I expect that to change, but during the period I was building — March to August 2026 — Claude gave better answers and better code.

There was also an odd pattern. For about a week after a new version shipped, it felt sharper; past that week, it got noticeably worse. It happened every time, so I don't think I imagined it.

And then I was banned

This was the decisive part. The game was originally built for children, and one day my account was banned for "possible child abuse". I appealed. It was never reinstated.

I have no idea what triggered it. Looking around, I found people building children's apps and educational worksheets who had been banned the same way. My guess is that development involving the word "children" tripped an automated classifier.

What struck me hardest was this: thank goodness everything was in GitHub. The account was gone, but the code, the reasoning behind every decision and the rules were all still there. I lost the conversation history and nothing else. Development didn't stop.

This is the most extreme illustration of article two. Information that exists only inside an AI disappears along with the account.

What I did: went back to Claude. And stopped changing tools in favour of changing process. If every AI produces accidents, the thing that prevents accidents isn't the AI — it's rules and verification. And staying in a shape where any AI becoming unavailable doesn't stop you. That turned out to be the best insurance of all.

Sometimes an AI just gets stupid

The same AI that produces results like this can suddenly lose the plot
The same AI that produces results like this can suddenly lose the plot

It's normally fine, and then comprehension falls off a cliff. It forgets what I explained a moment ago, misreads simple instructions, repeats the same mistake. I call this "the Nabeatsu phenomenon", after a Japanese comedian whose act is to act like an idiot on every multiple of three. Normal most of the time, suddenly broken under specific conditions.

It happens most when the conversation is long, when several topics are mixed together, or when my instructions were vague.

What I did:

Why I keep going anyway

Having complained at length: without AI, this game wouldn't exist. The physics engine, processing the artwork for all the items, the hazards on each planet, the site, the drafts of these articles. It's simply too much for one person in their spare time.

For people who like making things, this is a dream setup

I've spent years working on the web and on games, for the simple reason that I like making things in digital form.

Working inside an organisation doesn't mean you get to build freely, and it has plenty of downsides. But collaborating with others, and building with a company's budget, were real advantages. Alone, you carry planning, design, implementation and art all by yourself. That's unrealistic, which is why the company structure was necessary.

That changed.

With AI, roughly three person-months of planner, director and programmer time fits into $200 a month. The first time that sank in, I was slightly stunned. Things that once required a budget, a team and a schedule before you could even begin now start moving in one person's hands.

You can build freely without belonging to an organisation. That's the biggest benefit, in my view. Build what you want, the way you want. No meetings, no sign-offs. A spec you think of at midnight can be tried out that same night.

Especially for people who have worked on websites and app services inside a company and like making things, I think a remarkable environment has arrived. Your domain knowledge, your ability to pin down a spec, your eye for quality all carry over directly. The missing piece — hands to do the work — is what the AI supplies.

None of which means it becomes effortless, as everything above should make clear. An AI is best thought of as a capable subordinate who never double-checks. There's a lot you can delegate. But the final check is always yours. On that footing, you can get further than you'd expect.

What I wish I'd known

This series is still going. I'll add more as I remember things, or run into new ones.

Series: "I wish I'd known this from the start"

← Back to the dev log