The stack in ten minutes

You need four things, and nothing else — no game engine, no asset store, no build step.

1. Claude Code running Opus 5.5 — an agent with a shell, so it can write files, run them, and look at the frames. The chat app can write a game; only an agent with a terminal can render it and critique it.
2. A renderer. For 2D, nothing but an HTML file and the canvas API (what Orbit Switch uses). For 3D, three.js copied into the project folder so there is no CDN dependency and the game works offline.
3. A headless browser (Playwright + Chromium) so the agent — and the trailer pipeline — can screenshot frames.
4. ffmpeg to turn frames into video.

# One project, offline-ready
mkdir orbit-switch && cd orbit-switch
# (3D only) a local copy of three.js so there is no CDN and it works offline:
#   curl -L -o three.module.js https://unpkg.com/three@0.160.0/build/three.module.js
pip install playwright && python -m playwright install chromium
# ffmpeg: brew install ffmpeg  (macOS)  |  apt install ffmpeg  (Linux)

# Run the agent on Opus 5.5
claude --model claude-opus-5-5
# set effort: medium for small fixes, xhigh for a new game, max for a flagship piece

Two notes specific to Opus 5.5. Its thinking is always on, so it plans before it writes — good for multi-step game code. And it is multimodal, which is the whole basis of the playtest loop later. Confirm your setup with a one-line smoke test: ask it to write a one-file canvas page that draws a bouncing dot with a window.render(t) function, render frame t=2.0 to a PNG with headless Chromium, and then open the PNG and describe it. If that round-trips — it can write code, run a browser, call ffmpeg, and see an image — your studio is ready, and every later step will work.

A cost note so there are no surprises: Opus 5.5 is $4 per million input tokens and $20 per million output (cache reads $0.20). A full small game is a handful of passes, so it is dollars, not hundreds — but set the effort deliberately rather than running everything at max.