B-TIER demo
Google game Playground — why a harness that can't see its output fails
100k ft — What is this?
Google game Playground — why a harness that can't see its output fails
50k ft — Why care?
A hands-on review of Google's prompt-to-game Playground: output ignored the supplied concept art, the model denied receiving it, and it admitted it cannot see or play the games it builds. The takeaway is that a game-building harness needs visual verification and playtesting.
10k ft — Key takeaways
• Agent harnesses for visual work need a way to see and test the output
• Paid tool, low quality per this test
• Without feedback loops, generated games ignore reference art
5k ft — Tools & techniques
Google Playground
visual verification loop
1k ft — Deep-dive / try it
I love Google, but what a flop.
Transcript excerpt
Google just killed game developers. They just dropped their playground, which lets you create custom games in minutes. I'm totally kidding, this product is trash. When you get in here, you see a bunch of different games that you can actually play, because you can upload your games that you make into this library and other people can play them. This is currently the game with the most plays and the most thumbs up, and it is pretty bad. Like look how this looks right now. This is something that I could have built with Haiku 3.5 two years ago. And look how funny this is. This is the concept art I gave it. This is the setting, these are the towers, these are the enemies, and look what it built for me. It built me this world, and these are the slime characters, these are the towers, here's the arrow tower, and here's the cannon. This looks absolutely nothing like the concept art that I gave it. And then it tried to gaslight me. It literally said, hey, you never uploaded a concept art. And then I said, yeah, I did, here it is again. And then it said, okay, here, let me try to fix it. But it still just looked awful. And what's super interesting is that there's no way for it to actually visually verify anything. Look at this, I said, so I just want you to confirm to me that your job is to build me games, but you can't look at the game or play the game yourself. And it said, yep, that's correct. All I can do is build the code and optimize the code, but I can't actually see anything…
More from @nateherkai
- A-TIER Claude Opus 5.5 to Blender — 3D animation built entirely as code
- A-TIER Claudex skill — Claude and GPT-6 Astra cross-audit plan and build
- A-TIER Claude trading watcher — LiquidTrade + QuiverQuant for insider-trade alerts
- A-TIER Claude design stack — Emil Kowalski, Impeccable, taste skills plus Figma MCP
- A-TIER MapQuest MCP for Claude — location-aware agents that route your day