I let Grok build a game, played it, and deleted all of it

Good docs and a cheaper builder didn’t give me a good game. Swapping the builder didn’t fix it either. I wrote the docs, handed them to Grok, played what it built and deleted every line of code. Claude rebuilt it from the same docs and I rejected that too.

I’ve wanted this setup since August: all the rules and skills written down, so a cheaper model can build things from them. Grok runs in Cursor.

This week I wanted a game built, so I wrote the whole set. Rules for what makes a good game, a stack of playbooks, skills, plan templates, and a launcher script that starts Grok on a task at its highest effort setting. One game got its own plan of 47 tasks. The deal is simple. I write the plan properly, Grok builds.

70 minutes, then gone

Grok worked through the first ten tasks on Monday. Baseline commit to last task took about 70 minutes. It was a multiplayer brawler, and I played it that afternoon.

ok this is huge piece of shit and grok will never do it, we need to clean absolutely everything that useless agent did and then we gonna do it with you from scratc in new session so we keep only docs

The next commit removed the Grok build and kept the design docs. I logged the decision: Claude builds the games, Grok doesn’t.

Claude’s rebuild was rejected too

Claude rebuilt the same ten tasks from the docs in about 17 minutes. Ten commits, check script green. I played it and it was no better. It looked like junk, felt bad to play, had broken stuff and no UI.

Green checks mean nothing here, because they never played the game. Claude built it blind. The session ran from the hub folder, so the game engine connection never loaded and Claude never saw the thing running. Next I’m starting Claude inside the project so it can play the game before I do.

I keep asking for Grok anyway

The decision didn’t hold. On Tuesday I started a new game and asked for the docs and architecture to be prepped, then multiple Grok agents sent in, with Claude building the interface and assets. I also have a theory about the first failure:

maybe just the idea itself was dog shit so just make sure it really spends attention and does not do fast lazy stuff and follows the docs

On Wednesday I was asking for “do all the code needed with grok”. Then I looked at a build and wrote “gameplay kinda dog shit, spawn is dog shit, no onboarding”, plus a map that needed terrain and a real city look. That one went to Codex.

Today, with Codex finished, I asked if we could build the whole game now and orchestrate Grok agents on Cursor in parallel, checking their work. I looked at what came back and it’s no good either. No map, nothing. Claude and Codex built that one, not Grok.

So that’s where it stands. Grok’s first build got deleted. Claude’s rebuild got rejected. A later game built by Claude and Codex is junk too. The docs existed every time, and I keep asking for Grok through all of it.

I don’t know what exactly was wrong with the first Grok build beyond that one sentence of mine. I kept the docs. None of the games is finished.