Vomit: Clean up Claude 5's token output with a separate LLM
Points and comments are a snapshot, not live.
Vomit cleans up Claude Opus 5's verbose, unnatural prose by piping output through a local LLM.
The tool, written in Go, intercepts Claude's markdown comments and has multiple commands including 'scrub' and 'tail'. It replaces token-heavy, anthropomorphizing language (e.g., "caveat is a real one") with plain English. The README and commit messages show the developer iterated on the tool with Claude's assistance, adding Windows compatibility and a command to rewrite pull request blurbs. The prompt instructs a second LLM to remove distracting beats, pseudo-epiphanies, self-praise, and subject-verb oddities, and to avoid em dashes. The project is fully local with no telemetry or external dependencies.
What commenters are saying
Commenters largely agree that Claude Opus 5's writing has degraded into verbose, unnatural "token vomit." One shared a blog post with a before-and-after example showing a Claudish message about a force push rewritten clearly by the tool. Several users noted similar projects like "claudish-to-english" with a shared prompt. A few commenters defended the originals as usable but acknowledged the volume of output makes it exhausting. One observed the prose reads like notes from a developer deep in detail. Another asked about the root cause, wondering if watermarking or over-training is to blame. Some recommended trying other models like Sol or Muse Glimmer 30B.