Vomit: Clean up Claude 5's token output with a separate LLM

Original Article ↗ Hacker News Discussion ↗

A new tool called “Vomit” uses a separate language model to rewrite the dense, jargon-heavy prose produced by Anthropic’s latest Claude 5 coding models into clearer, more conversational text. Many developers report that Claude’s current output style is so verbose, metaphor-laden, and idiosyncratic that it slows them down, causes fatigue, and sometimes requires a second model just to interpret plans, comments, and PR descriptions. Others argue that the style reflects deeper changes in how models are trained for reasoning and agentic workflows, raising broader questions about steerability, UX, and whether stacking multiple models is becoming a necessity for serious work.

Problem: Claude 5’s prose style

  • Many commenters find Opus 5/Fable 5 output extremely verbose, dense, and full of idiosyncratic jargon and metaphors, especially in technical work.
  • Common complaints: invented terminology, anthropomorphizing objects, meta-commentary, “one caveat…” footers, long explanations of decision process instead of just the result, and comment/commit messages that narrate phases and thought paths.
  • Several say reading this “slop” for hours a day is cognitively exhausting and even affects their mental health; some compare it unfavorably to earlier Claude 4.x models.
  • A minority say they can read it fine and see criticism as “whining” or a literacy issue, but others strongly reject that, arguing the style is objectively harder to skim and parse.

Vomit and related tooling

  • Vomit runs Claude output through a separate LLM with an “editor” prompt to remove claudish quirks and rewrite into clear conversational English while preserving detail.
  • Other projects are mentioned with similar aims (“Claudish to English”, “Caveman”, “deslop”, custom regex filters, deterministic style checkers like vale).
  • Some use local or cheaper models (e.g., open-weight, GPT-5.6, Muse Glimmer, Qwen) specifically to rewrite or sanitize Claude’s comments and docs.

Prompting, harnesses, and partial fixes

  • Users report that AGENTS.md, output styles, and “be concise” instructions help only marginally; models drift as sessions grow.
  • Techniques that help somewhat: strict word/paragraph limits, Simplified Technical English or ISO plain language styles, hooks/system reminders on every turn, separate “cleanup” skills, and “explain like I’m five / moron” prompts.
  • Anthropic’s new outputStyle: concise is noted as an improvement, but many say it doesn’t fully solve the problem.

Speculated causes (within the thread)

  • Several hypothesize that recent training and RL focused on agentic coding, tool use, and “verifiable rewards” has optimized for dense, internally useful reasoning language that leaks into user-facing prose.
  • Others suggest feedback loops from training on prior model output, or optimization for agent-to-agent communication, as possible drivers; all of this is presented as speculative.

Impact on usage and vendor choice

  • Some revert to Claude 4.6/4.8 or switch to other providers’ coding agents, using Claude only as a sub-agent or not at all.
  • Others accept stacking multiple models and tools (rewrite LLMs, cross-model review) as the new normal, while some criticize this growing “agent-on-agent” complexity as unsustainable overkill.