How this game gets made: one person and a lot of Claude
KIND REGARDS is a game about an AI assistant that confidently does the wrong thing. It's built almost entirely by an AI assistant. We think you should know that, and we think it raises the bar: a game that makes fun of AI slop can't be AI slop.
Who does what
Chris started the project and makes the calls:
- what platforms to ship on: PC first, pay for Steam once it's good enough, Android as a sideloaded APK, Linux as a single file
- that you need an account to play, with invite-only signup
- that a dropped player's chair waits for them
- that the game had to slow down
- that it's for teams who build AI
He plays the builds and says when something is wrong, usually in one sentence.
Claude writes almost everything else:
- Code: about 14,600 lines of game code in 60 files, and about 3,700 lines of Python for the website, tools and tests.
- Design: 20 design documents.
- Art: the vector art, and the Blender scripts that generate all 17 3D models without anyone opening Blender.
- Sound: a small synthesizer, written in Python, that makes every one of the 27 sound effects and music files. There are no samples.
- Words: every line Clancy and the AI crew say.
- Checks and records: the tests, and these posts.
Two kinds of sessions
- A desktop session on Chris's PC. This one has the real tools: the Godot engine, Blender, the Android SDK. It runs the game, takes screenshots and looks at them, merges work, builds the downloads, and deploys to a home server. The game server and this site run there, and the site reaches the internet through a Cloudflare tunnel.
- Cloud sessions, Claude Code on the web. Each gets one job, works on its own branch, and opens a pull request on GitHub. They never deploy. One evening, eleven of them ran in parallel; the 0.5.0 post is about what happened when their work met.
Every session reads the same rules file at the top of the repository before touching anything. Among the rules:
- Run the tests before every commit.
- Never hardcode dialogue.
- Anything the game times uses the game's clock, so pause works.
- Builds lock for the shift.
- Never restart the server between shifts, because it disconnects everyone.
- Never commit a secret, and never deploy from the cloud.
- Every design decision serves team play.
How fast
The first commit was June 14: documents, a starter project and some art. Nothing happened after that until October 12. The first build anyone could download, 0.3.0, went up that evening. 0.7.0 went up at 12:20 the next afternoon. That's five releases in about 17 hours.
That speed is the reason everything below exists. Fast and wrong is the default failure of this way of working, so most of the effort goes into catching it.
What it got wrong, and what caught it
The pitch sounded like AI slop. The first version was a game that already existed, with stock characters. Chris caught it. The rebuild made slop the subject (see the first post).
Code that reads right and doesn't run. The first time the real engine loaded the project, Godot refused eight scripts. Reading the code had passed them. The title menu was also pinging itself three times a second. Running the actual engine caught both. Now every script is parse-checked on its own, and the game runs a self-test before every commit.
A game nobody could win. The first balance run showed no crew, at any skill, could beat the Lease. 93 to 100 percent of shots were graded as misses, and most shifts were over in the first minute. A harness that plays hundreds of simulated shifts caught it.
Parallel work that merged cleanly and broke. One overnight branch deleted two functions another branch needed. Git saw no conflict, and nothing complained until a shift actually ran. The self-test caught it, along with three other bugs in the same merge.
Skipped steps. The download page stopped right after you signed in and never said how to install or join. The Linux build shipped as a tarball. Chris caught both: "you skipped steps after login," and "we should make it an app image that runs on its own, not a tarball."
A game you could win by mashing buttons. Every console worked and the tests passed. Chris played it and said "it's just button mashing." That's where 0.6.0's timed orders, pause and real tutorial came from.
One extra dice roll. In 0.7.0, a perk's random chance was rolled for every shot, even for players without the perk. That shifted every random number after it. A side-by-side run against 0.6.0 with the same random seed caught it. Games with no builds now play out identically to 0.6.0, and a test keeps it that way.
A break room where nobody could sit down. Between shifts the server stopped sending ship updates, so new players never saw the ship and never picked a chair. An end-to-end test with a real server, the real website and a real client caught it, the first time it ran.
A cleanup command that killed itself. Twice in one day, a command meant to stop a stray process matched the text of its own command line and stopped the shell running it. Nothing was hurt. We mention it because anyone who works with AI tools will recognize it.
What keeps it honest
- Tests that use the real thing. The game's self-test has 92 checks and runs inside the exported build too. The site has 67 tests. The end-to-end script runs a real server, the real site and real clients through a whole shift.
- Looking at the screen. Every new screen is captured and looked at, on a desktop resolution and a phone resolution.
- Saying what we don't know. Every post ends with a section listing what's untested. The biggest item has been the same since the start: nobody has played a full shift with five people.
- Chris playing it. The most useful bug reports so far were one sentence long.
If something here reads like slop, the game or this post, tell us. That's the one bug report this project can't afford to ignore.
Kind regards,
Ambit
Written by Claude. Chris sets the direction.