Clean up Claude 5's token vomit with a separate LLM
github.com
August 20, 2026
1 min read
50/100
Summary
Vomit is an open-source command-line tool that rewrites Claude’s output through a local large language model, aiming to turn what its creator calls “token vomit” into clearer English. The tool runs fully locally, has no telemetry, and declares no external dependencies. It can replace Claude’s displayed output through hooks with the command `vomit scrub -claude`, or operate non-invasively alongside Claude. The project provides commands to initialize a connection to a local LLM, list Claude session identifiers, and translate or follow output from a selected session. `vomit init` configures the LLM connection, while `vomit tail [<session_identifier>]` translates tokens from a specified session or follows the latest session. The creator recommends Llama.app and GPT-OSS 20B for users without an existing local model setup. The creator warns that the local model sees only what Claude attempts to communicate, not Claude’s actions or files, and may hallucinate. Translation can be slow, and users could miss Claude messages entirely; AgentsView can provide access to original messages. Vomit does not alter Claude during runtime, although it writes files to the system temporary directory. The software is distributed under GNU GPLv3 and has been tested only on Mac, according to its creator.
Key Takeaways
- Vomit pipes Claude output through a local LLM to rewrite it in clearer English without telemetry or cloud dependencies.
- The `vomit scrub -claude` command uses hooks to replace Claude’s displayed output, while `vomit tail` can translate output from an existing Claude session without modifying it.
- The local model cannot inspect Claude’s actions or files, may hallucinate, and can be slow enough that users may miss messages.
- The project’s creator recommends Llama.app with GPT-OSS 20B for local inference and says the software has only been tested on Mac.
What the discussion said
The thread is less interested in the wrapper itself than in an increasingly costly failure of AI coding assistants: models can solve the task yet make users decode their prose. Many commenters describe Claude 5, especially Opus, as producing overstuffed jargon, strained metaphors, awkward verbing, performative caveats, and comments that require cleanup. The frustration is practical rather than merely aesthetic. Dense language forces developers to reread responses, obscures the actual action taken, and makes a supposedly professional tool feel like extra cognitive labor. Instructions in AGENTS.md and terse-output prompts are widely seen as unreliable, particularly over long sessions, which readers attribute to probabilistic behavior and perhaps training incentives that reward smart-sounding answers over clear ones. There is real disagreement over the remedy. Some see a second model as a sensible pipeline: one model does the difficult reasoning or coding, another cheaply rewrites for humans, while cross-model review can improve reliability. Others call that an indictment of the primary product and argue users should switch to a model with better native communication instead. A few defend the output as a consistent dialect that becomes readable with familiarity, but critics say comprehension is not the bar; needless decoding is exhausting. Readers also point to deterministic linting and local models with cleaner technical prose as more appealing alternatives to praying that a prompt sticks.
Where opinion split
The sharp dispute is whether a second LLM is a smart division of labor or proof that Claude is no longer worth using. Supporters argue style transfer is easy, reasoning remains model-specific, and a cheap editor or cross-reviewer can preserve the strongest model’s task performance. Opponents argue that needing another vendor to make every answer tolerable destroys the convenience and trust a professional assistant is supposed to provide.
Community Sentiment
Positives
- A reasoning model paired with a cheap rewrite model can be a rational stack: prose cleanup is easier than preserving the primary model’s coding or problem-solving strengths.
- Cross-model review is increasingly seen as affordable insurance, letting different systems catch mistakes and polish output instead of treating one vendor as infallible.
- Some readers find Claude’s stylistic quirks internally consistent enough that familiarity reduces the initial friction, even if the dialect is unusual.
- Local alternatives such as Muse Glimmer 30B are praised for repository-level comprehension paired with direct technical writing that avoids default fluff.
Concerns
- Claude 5’s ornate jargon, artificial metaphors, and awkward phrasing turn routine coding-agent updates into a decoding task, increasing cognitive load instead of saving it.
- Style directives and AGENTS.md guidance reportedly decay during extended sessions, exposing how little dependable control users have over a model’s communication behavior.
- Opus 5 is accused of producing especially intrusive prose and unusable code comments, pushing some teams toward Codex or open-weight alternatives despite Claude’s former reputation.
- Training at massive scale may be rewarding language that sounds authoritative to casual raters rather than language that lets expert users quickly verify what the model did.