Make your coding agent play its own game before it says done

Side-by-side card from the author's video: "Claude Code vs Grok Bot", the shared racing-game prompt and the Claude Code neon track

@theSethian gave Claude Code and Grok Bot the same prompt: a 3D browser racing game with a neon track, an AI opponent and a 3-lap timer, then race once to prove it works. Per the author, Claude Code finished in 1:43 with a scripted driver that moved the car through code. Grok Bot took about 16 minutes because it drove the game with its own controls on its cloud computer, found the steering was reversed, fixed it and finished all 3 laps. The author's tip: add "play it yourself before you say done" to agent prompts.

Try it yourself

What you get

A build the agent has actually used through its own controls, so input bugs like reversed steering get caught before you see them.

Tools

Time / cost: Author: Claude Code 1:43; Grok Bot about 16 minutes including its test race

Steps

  1. Give the build spec and end it with an explicit test: open it and play one full round.
  2. Require the test to use the same inputs a person would (keys, clicks, touch).
  3. Ask the agent to list every bug it hit while testing and how it fixed each one.
  4. Compare against a second agent or your own play-through if the bug would be costly.

Copy-ready prompt

Build a playable 3D browser racing game in a single HTML file: neon track with curves, AI opponent car, lap timer, 3 laps, finish screen with your time. Open it and race one full race to prove it works.

Prompt quoted verbatim from @theSethian’s public post. Check the output yourself and approve before anything is sent or bought.

Watch out: One informal run by one person. The author says Claude's steering was right on the first try; a scripted test can still miss input bugs, and a played test takes longer.

Proof / source

Self-reported Author's X post with a side-by-side replay video and the prompt on screen (Oct 14, 2026). Self-reported.

In their words

Claude Code wrote the game, let a scripted driver finish it and reported back in 1 minute 43 seconds. Grok Bot also wrote a version, then raced it on its own cloud computer. Its first test race showed the steering was reversed, so it fixed the bug and went back for all 3 laps.

— @theSethian on X

Read the original source ↗

Added · How we verify · Report an error

Related