I did not start with code. I started with a long message that said, more or less: let us fine grain the plan before you begin. Set up the rules for the assistants first. Use the Claude Code CLI on this PC for the big planning model, because I am paying for the subscription anyway. Add a setup phase that downloads the tools, and test that the VR headset works before anything else.
A game built by a machine that forgets
It is a strange way to begin a game. No world, no character, only a file that tells future sessions how to behave. But I had a feeling about this project from the first hour: it would be built by many short conversations with a machine that forgets everything between them. If the rules were not written down, I would lose the thread by Sunday.
By the afternoon I had put some keys in a file and asked the question I would ask many times in the next ten days:
Be honest, do I need to switch to a new session with you?
The honest answer is usually yes. A fresh session thinks better than a tired one. I asked it to check that no stray API key sat in my shell profile, because a key there would quietly make the planner bill me instead of using my subscription. It was clean. I wrote "I think I am ok then" and told it to run the first three setup steps.
Tools on the bench
By evening the repo had a skeleton, a check against committing secrets, and every tool pinned to an exact version: the game engine, the VR toolkit, the streaming software for the headset. A doctor script checked that the whole chain was in place. It felt like laying out tools on a bench before building a table.
Then, close to midnight, the other question: "what if I want to try Wi-Fi?" I had a Quest 3 on the desk and a cable I did not want to use. That question belongs to the next day.