A game in development

What it is

An action RPG in Unreal Engine 5.8. It's in development: there's a playable packaged build, I'm playtesting it on a controller, and it is not shipped. Right now the work is cleanup, truthful tests and validated data, not new content. It stays unnamed here until it's closer to done.

Why it's on this site

Because the infrastructure turned out to be the interesting part. I measured what's underneath the game on 20 September, and the numbers say more about how an agent-driven build actually runs than any finished feature would.

The claim isn't "AI built a game." It's how much tested engineering a few hundred prompts of direction can carry. The agent writes the code, the tests and the documentation. I direct, decide, buy things and playtest.

How it's built

I ran 26 orchestrating Claude Code sessions. They spawned 519 subagent conversations, which out-produced the main thread 7.8 to 1 by output volume.

My side of it: 157 prompts, about 53,000 typed characters of direction, in the two weeks when three-quarters of that C++ was written. The other quarter came from a two-day scaffold in July whose transcripts didn't survive, so I can't count those prompts. I checked that proportion with git blame rather than guessing.

The documentation is bigger than the code: 104,912 lines of Markdown against 81,707 lines of gameplay C++, 1.28 lines of docs per line of code. The largest share of it (about 40%) exists so the agent can read itself back into context at the start of a session. Most of this project's documentation has no human reader at all.

What the agent couldn't do:

• 250 recorded product decisions
• purchasing and licensing calls
• installing the engine
• adopting a design document wholesale
• playtesting with a controller in hand

What broke

Nine weeks of nothing. The commit history shows two days of scaffolding in July, then 63 days with zero commits, then 257 commits in two weeks. From the outside that looks abandoned. It wasn't: the engine version I was on lacked a capability the project needed, and work resumed when the next version shipped.

Tests that couldn't see the game. 30,069 lines of automated tests, and the defects that mattered most still showed up with a controller in my hands.

My own stats report drifted. The first draft of the measurement report had understated commits, decision entries and subagent counts, and a pattern-matching bug counted blank answers in my decision inbox as answered. A fresh-context check and a second pass caught both before anything was published.

Receipts

All measured. More on the receipts page.