Dev post-trains a VLM to catch game bugs Claude misses: 56% to 81%
OdinLovis · x · 2026-11-18
Developer OdinLovis notes Claude can't truly watch video—it samples screenshots every half second, so one-frame glitches and animation bugs slip through, making game debugging hard.
His fix: post-train a dedicated VLM on a large dataset of video games originally built with Claude. The VLM watches every frame, tells Claude what broke and when, and learns Claude's typical mistakes.
Results:
- Claude alone: 56% of bugs caught
- Claude + his VLM: 81%
- Fast glitches (flash, stutter, camera jump): 30% → 100%
The project ships soon. He argues post-training remains the key to pushing further even when working with closed-source models.
More from coding & agent
- Dev teases faster, cheaper game creation with his bug-catching VLM stack — OdinLovis · 2026-11-18
- Pipecat open-sources Audience: a tiny classifier so voice agents know when they're being addressed — alexcovo_eth · 2026-11-18
- Giving Your Agent Your Own API Token: An Infra Engineer's Cautionary Tale — WesternNo2429 · 2026-11-18
- Weave: Model Writes One Script, Rust Runtime Parallelizes Independent Tool Calls — Business_Rain_1042 · 2026-11-18
- Microsoft open-sources Work IQ Starter Pack to fill demo tenants with realistic data — DanWahlin · 2026-11-18
- GitHub publishes harness-level data comparing Copilot CLI vs Codex and Claude Code — 0xkarasy · 2026-11-18