Vomit: Cleaning Up Claude 5 Token Output with Local LLMs

Overview

Vomit is an open-source utility designed to intercept and rewrite the output of Claude 5, which users have described as "token vomit"—prose that is overly verbose, stilted, and occasionally incomprehensible. The tool pipes Claude's output through a separate, local LLM to translate the "Claudish" style into clear, conversational English.

Built in Go, Vomit is designed to be fully local with no telemetry and no external dependencies, ensuring that the processing of the rewritten text remains on the user's machine.

Core Functionality and Installation

Vomit operates by utilizing LLM hooks to replace Claude's output in real-time or by running in a non-invasive mode to translate session logs.

Installation and Setup

To install and configure Vomit, users follow these steps:

  1. Install the binary: go install github.com/zachahn/vomit@latest
  2. Initialize connection: Run vomit init to set up connection details for the local LLM.
  3. Configure hooks: Run vomit scrub -claude to set up the instructions for replacing Claude's output.

Usage Modes

  • Invasive Mode: Replaces Claude's output directly via hooks.
  • Non-Invasive Mode: Allows users to translate sessions on the side using commands such as vomit list (to see session identifiers) and vomit tail (to translate a specific or the most recent session).

Supported Local LLM Backends

Vomit is compatible with any LLM provider that uses the OpenAI API, including:

  • Llama.app (Recommended model: GPT-OSS 20B)
  • Ollama

Technical Implementation: The "Claudish" Translation Prompt

The effectiveness of Vomit relies on a specific system prompt used by the local LLM to act as an editor. The prompt explicitly instructs the local model to identify and remove specific "strange characteristics" common in Claude 5's output:

  • Roundabout reasoning: Removing pseudo-epiphanies and distracting beats in the flow.
  • Incorrect Subject-Verb combinations: Ensuring that only humans, groups, and agents perform "action verbs," while objects remain passive.
  • Specific banned words: The prompt targets words like "carries" and "names" when used as action verbs for objects.
  • Stylistic distractions: The removal of em dashes (—) to prevent distracting beats in the prose.

Community Analysis of Claude 5's Prose

The creation of Vomit has sparked a wider discussion among developers regarding the degradation of prose quality in recent Anthropic models, specifically Opus 5.

The "Claudish" Phenomenon

Users report that Claude 5's output has become increasingly difficult to read, characterized by:

  • Deliberate obfuscation: The use of dense jargon and stilted metaphors to sound more authoritative.
  • Agent-centric optimization: Some speculate that the model has been RL-optimized for agent-to-agent communication rather than human-to-agent communication, leading to bizarre word choices and excessive self-justification.
  • Cognitive burden: Developers have noted that the verbosity and "Columbo-style" footers (adding caveats at the end of every response) increase the mental effort required to parse technical information.

Alternative Workarounds

Beyond using Vomit, the community has suggested several other methods to handle "slop":

  • Prompt Engineering: Using incantations like "You must use ASD-STE100 Simplified Technical English (STE)" or using tools like "Caveman" to force a simpler style.
  • Model Switching: Moving to Codex or open-weight models (e.g., Muse Glimmer 30B) for technical writing.
  • Version Downgrading: Some users report that Opus 4.6 was the last version without these prose issues.
  • Official Settings: Recent updates to Claude Code have introduced an outputStyle: concise option to mitigate verbosity.

Theoretical Causes

Some contributors suggest that the strange prose is a byproduct of the model's reasoning process. The theory is that the "strange" reasoning prose is effective for the model's internal looping and systematic thinking—which has improved coding and math performance—but fails to translate into a deliverable, human-readable answer without a final "humanizing" stage.

Sources

Related

  • Dispatch
    Claude Opus 5.0 language regression: incoherence and toxic verbosity

    Claude Opus 5.0 (and 4.8) exhibit a regression toward verbose, jargon‑heavy, and often nonsensical output that users describe as toxic and unreadable, prompting workarounds and calls for a plain‑language register.

  • Dispatch
    nobuzz: Removing 'Claudisms' from Claude Code with Gemini

    nobuzz is a Claude Code skill that uses the Gemini CLI to translate Claude's verbose, clickbait-style responses into plain, professional English.

  • Dispatch
    Removing Claude AI Claudisms: Using MessageDisplay Hooks for Vocabulary Control

    Learn how to use the MessageDisplay hook in Claude to programmatically replace annoying AI-specific phrases (Claudisms) like 'load-bearing' and 'honest take' with custom text.

  • Project
    dongbeixiaohuo/writing-agent

    Writing‑Agent is an open‑source, stage‑driven AI writing pipeline (CLI + optional Windows app) that structures article creation into discrete, auditable steps—topic, stance, evidence ledger, outline, opening hook, draft, human‑izer, fact‑checker, reviewer, reader test, and export. It works with DeepSeek‑V3.2, 智谱 GLM, or MiniMax via Anthropic‑compatible APIs, stores SHA‑256 hashes for reproducibility, and can output plain‑text and WeChat‑style HTML. Install via the Windows preview installer or clone the repo, run `npm ci`, and launch Claude Code. The repo includes a full demo showing every intermediate file.

  • Project
    KinetiNode/claude-fable-5-system-prompt-clean

    Vendor‑neutral Claude Fable 5 system prompt adapted for use across multiple frontier LLMs.