Did any of the AI's mistakes crash the game or throw an error?

No. Over the month, not a single bug caused a crash or an exception. Every mistake appeared as code that kept running without an error.

What kind of mistake was most common?

Features that existed in the code but were connected to nothing else — 15 cases, the largest of the 10 types. The example: wood never increased after 100 days because the button to select the action on screen was missing.

Were the mistakes about Unity itself or about the developer's own game?

Almost all were about rules unique to this game. Only one mistake in the month was about how Unity works, and one was caused by something outside the engine.

Large Language ModelsAI Coding AssistantsZenn AI/ML20h agoOct 8, 2026, 22:00 JST

Claude built a Unity game: 88件 of mistakes, 0 crashes

Send on LINE

3 Key Points

What happened
A solo developer who let Claude handle nearly all design and coding of a Unity game logged 88 mistakes in one month, against 138 tasks Claude finished. The largest group, 15 cases, was features that existed in code but were wired to nothing.
Why it matters
Almost all of those mistakes ran without any error, so they only surfaced when the game was played or measured — a failure mode that inspection or code review misses, and the developer notes the one Unity-usage error in the set was a single case.

WHO IT HITSSolo developers and small teams shipping games or apps on tight budgets, who cannot rely on crash reports or error logs to catch AI-generated mistakes, are the ones this lands on. Teams that review AI output only by reading code, rather than running and measuring it, may miss nearly all of the failures recorded here.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Get news like this every morning, free

Context & Analysis

The developer's setup is deliberately lopsided: he writes a full design first, then lets Claude implement it. That habit explains the dominant failure type he recorded. Claude implements what is written, but does not add the unwritten "connectors" between systems — and when Claude reviews the code, everything looks implemented, so it finds nothing.

The way the mistakes appeared is the more unusual part. Instead of surfacing as crashes or exceptions, they stayed silent and only became visible through play or measurement. That gap is what pushed the developer to change how he works: add a line to each feature spec saying where it breaks when it does not work, record how each mistake was caught, and spend review time on the rules unique to his own game rather than on Unity usage. His takeaway, after laying out all 88 cases, is that a day of playing surfaces what a month of reading code does not.

FAQ

Did any of the AI's mistakes crash the game or throw an error?: No. Over the month, not a single bug caused a crash or an exception. Every mistake appeared as code that kept running without an error.
What kind of mistake was most common?: Features that existed in the code but were connected to nothing else — 15 cases, the largest of the 10 types. The example: wood never increased after 100 days because the button to select the action on screen was missing.
Were the mistakes about Unity itself or about the developer's own game?: Almost all were about rules unique to this game. Only one mistake in the month was about how Unity works, and one was caused by something outside the engine.

Zenn AI/MLRead Original Article

Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Get it free

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.