Make a kart racing game with AI: how we built Capybara Kart

The brief is the most important file. Ours is 79 lines: the game, the controls, the online rules, the art style, the size and speed limits for phones, and the checks that prove it works. A short excerpt:

- First visit at most 1.3 MB over the wire, at most 750 kB before the first screen shows. Measure and report.
- Frame time on Chromium Pixel 7 with CPU throttling 4x in the busiest moment: report median and p95; target p95 at most 17.5 ms.

Numbers matter more than adjectives. "Fast on phones" gives an agent nothing to check; "p95 frame time at most 17.5 ms" does. Try this prompt to start yours:

I want to make a 3D kart racing game for phone browsers: <your idea in two sentences>. Interview me with up to 10 short questions, then write docs/brief.md: the game, the controls, the art style, online play, size and speed limits with numbers, and the checks that prove each part works.

We gave the brief to Claude Code with Claude Opus and asked for a plan and tickets. The plan set out the tools (three.js for 3D, plain JavaScript with no build step, a Cloudflare Worker for online rooms) and listed the risks, each with the ticket that reduces it:

Client and server rules drift apart: one copy of the rules in game/src/sim/, imported by the server (tickets 3, 7).

Each ticket has a goal, the files to create, the numbers to hit and a "done means" list. The order matters: the empty project and its test command first, one Blender picture of the game camera second, the race rules third, and the real art ninth.

Read docs/brief.md. Write docs/plan.md and then 8 to 12 tickets. Each ticket: the goal, the files, the numbers, and a "done means" list with a test for each item. Put the rules in plain modules with tests before any drawing. Do not write game code yet.

Ticket 3 wrote the whole race with no graphics at all: driving, drifting, laps, CPU karts. Its goal, word for word:

All race rules as plain modules that run in Node with no browser. The renderer later only draws what these modules say, and the server imports the same files.

That one choice saved us later. The rules have tests that run in seconds, a script races six CPU karts with no screen, and the online server uses the same rule files, which keeps online and offline races in step. Timing and the network can still make an online race feel different. OpenAI Codex coding agents built the tickets one at a time, each as its own pull request. The first ticket merged at 05:32 and the eleventh at 09:06: 3 hours 34 minutes.

Ticket 2 came before any game code: one picture of a street, rendered from the game camera, as a target for the look. Then ticket 9 built all 23 models (the capybara, the kart, hats, stalls, tuk-tuks, the bridge) from Python scripts that run in Blender with no clicking. Because the art is code, an agent can change it, rebuild it and check the file size the same way it changes the game. All the models together weigh 373 kB, under the 450 kB limit in the brief.

Here is how an online race works, in plain words. Each phone drives its own kart, so steering feels instant, and sends its position 10 times a second. A small server on Cloudflare runs one "room" per race. The room is the referee: it decides who got an item box first, who got hit, the laps and the finish order, and it drives the CPU karts. It sends everyone the whole race 15 times a second, and each phone draws the other karts about a tenth of a second in the past, so they move smoothly between updates.

The brief set hard limits so it fits Cloudflare's free plan, where 20 incoming messages count as one request. The server's fake-clock test of a full 3-minute race with 4 players measured 7,239 incoming messages ÷ 20 + 5 connections + 4 match requests = 370.95 billed requests. The Workers Free plan allows 100,000 requests a day (as of November 2026). That count covers only the requests the race room bills. The room also uses Durable Objects run time and storage, which have their own free limits on the Durable Objects pricing page, and each visit to the game page is a separate request. After a few real races, check the usage in your Cloudflare dashboard before you count on the free plan.

Every ticket had to pass npm run check: 65 game tests, 39 server tests, a bot that drives a lap by touch on an emulated phone, two bots that race each other online, and a balance check where a clean bot must win under 60% of 40 races. Bots are good at finding crashes. They are bad at noticing a game feels wrong, which is why the next step exists.

Jay played the first build and wrote back two lines.

"capybara kart controls are reversed". Pressing left turned the kart right, on the keyboard and by touch. Every test passed, because the CPU karts and the test bot used the same wrong sign. The fix flipped the sign; the useful part was the new test, which presses the real keys and checks through the real camera that the kart's nose turns to the screen left.

"i don't think my position in the pack is accurate". Half a second after GO, the standings read 2, 3, 4, 5, 1, 0: the back rows ranked ahead of the front row. The brief for the fix told the agent exactly how to rank:

Rule: rank by (laps completed, then distance along the centre line on the current lap, then time at which that distance was reached). Project each kart onto the nearest centre-line segment, with a continuity check so a kart near a hairpin cannot jump to the far side of the track.

The new tests script races where the true order is known: an overtake in a corner, the finish line, a kart going backwards. Before the fix, the first two had 5 and 181 wrong frames. Both fixes were done in Claude Code, along with a redesign of the menus.