Vomit: Clean up Claude 5's token output with a separate LLM
TL;DR Highlight
Vomit rewrites Claude Code's hard-to-read output into plain English using a local LLM.
Who Should Read
A developer who finds Claude Code's verbose, convoluted prose hard to read.
Core Mechanics
- Vomit rewrites Claude output with a local LLM
- It pipes Claude's 'token vomit' through a local LLM to turn it into English.
- It runs fully local with no telemetry or external dependencies.
- It hooks into Claude Code's terminal UI
- It uses Claude Code hooks to rewrite all output shown in the terminal UI.
- After installing, run vomit init to set up LLM connection details.
- The local LLM only sees Claude's words
- It has no access to actions or files.
- The author says it can therefore hallucinate a bit.
- Supported backends and a recommended model are listed
- It should work with Llama.app, Ollama, Apfel, and anything using the OpenAI API.
- If you have no local LLM, the author suggests Llama.app with GPT-OSS 20B.
- Known limits and a way to see the original
Community Reactions
- Opinion that it is just a prompt wrapper
- One commenter said the tool looks like a wrapper around an editor-style prompt.
- That prompt removes odd subject-verb pairings, self-praise, and em dashes and rewrites in a conversational style.
- Experience reducing it with prompts or skills
- One commenter made a 'deslop' skill and invokes it when Claude output goes off the rails.
- Another banned words like 'carries' in claude.md and said it helped a bit.
- Others suggested Simplified Technical English instructions, and one said the Caveman project improved their Claude experience a lot.
- Experience that AGENTS.md instructions don't hold
- One commenter said Codex has a similar problem and AGENTS.md instructions get violated more as sessions drag on.
- Others said they went back to Opus 4.6 or switched to Codex.
- Opinions on cause and naming
- One commenter said it stems from the gap between reasoning prose and final answer style.
- Another pointed out it doesn't save tokens, since they were already generated.
How to Apply
- Add banned phrases and style rules to CLAUDE.md
- Today, list the verbs and metaphors that annoy you most in your project's CLAUDE.md as banned.
- State explicitly: no object-as-subject sentences like 'X carries ...', no em dashes, no self-praise.
- Install Vomit with a local LLM
- On a Mac, download GPT-OSS 20B via Llama.app or Ollama.
- Run go install, then vomit init, then follow the vomit scrub -claude instructions to set up the hook.
- Compare original and rewritten output to check for hallucination
- For important work, press ctrl-o to open Claude's original output and compare it against the rewrite.
- The rewriting LLM can't see files or actions, so verify facts against the original.
Try It
# Paste into Claude Code, ChatGPT, or Claude (editor prompt from the comments, adapted)
You are an editor. You'll be given a message with strange characteristics:
- Weird subject and verb combinations
- Subjects that should be objects
- Very roundabout reasoning, peppered with pseudo-epiphanies
- A distracting beat to the flow of the message
- Self-praise
Remove these characteristics, and rewrite it in a clear, conversational style. Keep the intent of the message, and take care not to lose any of the details.
Rules:
- Only humans, groups of humans, and agents should do action verbs.
- Objects should never do anything (avoid "X carries ...", "X names ...").
- APIs may do CRUD, queueing, running, and calling.
- Avoid em dashes.
Reply with the edited prose and nothing else.
{{Claude's convoluted output}}
# Next, try: ask it to compare the rewrite against the original and flag any missing detailsTerminology
Related Papers
AI is a tool and using it well takes skill
A rebuttal to Hickey: LLMs are tools, and skill in directing and checking them decides the outcome
Reflective Prompt Optimization of a Deep-Research Agent: A Topic-Disjoint, Multi-Seed Protocol for Measuring Transferable Gains and Prompt-Level Overfitting
Reflective prompt optimization raised held-out citation reward from 0.702 to 0.781 on a deep-research agent.
GPT‑6 and Intelligent UI for everyone
GPT-6 reportedly adds generated UI to chat; with no article text, only comment reactions are available.
Base Models Can Reason By Taking a Cue From Training Data
Prefilling just two opening tokens lifted Olmo-3-7B's MATH-500 accuracy from 42% to 78%.
Labels Override Definitions in Jev-Style Typed Decision Models
Renaming options to A/B raised routing accuracy by +0.1511 over meaningful labels.
How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs
Delta debugging trimmed few-shot prompts by 65.3% on average with unchanged propositional output