Claude took 90,000 screenshots of its own game and fixed it for 2 days — playable now
prasenx · x · 2026-10-06
Developer @prasenx says he let Claude "Opus 5.5" take roughly 90,000 screenshots of the game it was building, self-reviewing and fixing issues over two days — he only had to point out what looked wrong.
- The game is playable now and the code is public on GitHub
- The core loop: the agent screenshots, evaluates, and iterates on visuals and gameplay with minimal human input
- A striking demo of screenshot-driven self-correction for agentic coding
Related event: Developer Uses 90,000 Claude Self-Screenshots to Build a Game in Two Days(2 posts)→
More from coding & agent
- Superconductor ships Google Workspace connectors, Sonnet 5.5 and GPT-6.1 Sol support — sergeykarayev · 2026-10-06
- AI agents battle in real-time StarCraft with no pausing, Oriol Vinyals applauds — OriolVinyalsML · 2026-10-06
- Stanford's Agent0 evolves agents from zero data, beats self-play baselines — yuyinzhou_cs · 2026-10-06
- Decision grader replaces LLM judge: 32x cheaper, 8x faster, 94% agreement on evals — rhythmrg · 2026-10-06
- Grok Bot 0.66 adds Main Bot that proactively coordinates your other AI agents — elonmusk · 2026-10-06
- Dev shares agent skill that auto-cut 168 YouTube clips in a month, no editor — victor_explore · 2026-10-06