Neon Serpent: a one-shot test of the whole loop

Neon Serpent — the game this post is about — is not a project I planned. It's a yardstick. Go ahead, try it:

Playable, right here. Open full-screen for a bigger board.

The prompt

The entire build was a single message, sent over Signal with no further instruction:

Make a snake game in HTML. Make it as full featured as possible in a single standalone HTML file. When you're done send it over

That's it. One prompt, no revisions. The agent (Hermes Agent driving Qwen3.8-27B) took the message, wrote the whole file, syntax-checked it, smoke-tested it in a real browser headless, and sent the result back. The artifact is 750 lines of standalone HTML — canvas rendering, keyboard and touch controls, pause, speed ramp, obstacles, combo scoring, high scores in localStorage — with no build step and no dependencies.

What it was actually testing

The interesting part wasn't the snake game. It was whether an agentic system could carry a vague, open-ended ask all the way to a finished, verified artifact with zero hand-holding:

Interpret the brief. "As full featured as possible" is a judgment call — what features count, in what order, how much is enough.
Deliver a working artifact, not a description. The file has to run in a browser. Not "here's the structure," not a plan.
Verify before handing it over. The agent ran a syntax check and a headless browser pass before sending, so "done" means "ran without errors," not "looks plausible."

Verdict

It passed. The game shipped exactly as sent — one prompt in, playable game out, no follow-ups. That's the baseline I want from an agent on bigger tasks: same discipline, more surface area.