A Go tool called vomit pipes Claude 5's output through a local 20B model to make it readable.

104 stars, GPLv3, installs via go install, hooks output with vomit scrub -claude, recommends GPT-OSS 20B on Llama.app. The README: "Save your tokens, Claude 5 is hopeless." GitHub The 266-point thread drew 260 comments largely agreeing that Claude 5 writes in dense jargon ("load-bearing", "seams", "smoke") and that prompt instructions to stop get ignored as sessions lengthen. I've hit this exact wall. The dissent in the thread is the sharper point though: paying two models to undo one model's house style doesn't fix the model.