Vomit, an open-source tool from developer zachahn, pipes Claude 5's token output through a separate LLM to clean it up. It hit the Hacker News front page on August 20 with 300 points and 291 comments — a near 1:1 comment-to-point ratio, which usually means a thread that argued more than it upvoted.
The premise travels further than the tool. Fixing how a frontier model writes is being handled not with a better prompt or a fine-tune but with a second model bolted on downstream: a second inference pass, a second bill, and a second failure mode, spent entirely on the first model's prose. That the project is named Vomit tells you the register the author was going for.
There is no coverage beyond the repository and the HN thread — no benchmarks, no writeup, no independent testing. Read the source before wiring it into anything.
Sources: GitHub, HN discussion
