A Go Tool Called 'Vomit' Pipes Claude 5's Output Through a Local 20B Model to Make It Readable, and 260 HN Comments Agreed With the Premise
GitHub / Hacker News·high signal
Zach Ahn's vomit (104 GitHub stars, GPLv3) installs via 'go install', hooks Claude's output with 'vomit scrub -claude', and rewrites it through a local model, with GPT-OSS 20B on Llama.app as the recommended default. The README's motivation is blunt: 'Save your tokens, Claude 5 is hopeless.' The 266-point thread drew 260 comments, most of them agreeing that Claude 5 writes in dense jargon ('load-bearing', 'seams', 'smoke') and that prompt instructions to stop get ignored as sessions lengthen. The dissent is the sharper point: paying two models to undo one model's house style does not fix the model.