You’ve asked an AI to write a function. It returns a paragraph of sycophantic nonsense before getting to the point. You’ve asked it to summarize an email. It gives you a verbose essay with false epiphanies. You’re not alone. This is the dirty secret of conversational AI: it’s terrible at being concise. AI’s default output isn’t a conversation — it’s a monologue with a sycophant.
Claude, GPT-4, Gemini — they all suffer from what one developer lovingly calls ‘token vomit.’ The very features that make them ‘conversational’ — the padding, the ‘Let me think about this,’ the roundabout reasoning — make them painful to use for real work. One developer, Zach Ahn, created a tool called ‘vomit’ that does exactly that: it pipes Claude’s output through a second, cheaper LLM whose only job is to strip the fluff. The result? Clean, concise, actionable text. The best AI output is not generated — it’s edited.
This is not a hack. This is a design pattern. The future of AI isn’t one omniscient model that does everything. It’s a pipeline of specialists: a worker model that generates raw material, and an editor model that translates it into human-friendly language. Think of it as ‘Claudish to English.’ One user on the thread says: ‘I like the “Claudish to English” name better.’ Another adds: ‘I’ve been using the pattern of using coding agents to orchestrate my CLI agents and it’s really good for these kinds of things. The vomit never makes it my way.’ The vomit never makes it my way — that’s the gold standard.
The implications are huge. If you’re using AI for coding, writing, or analysis, you’re probably wasting 30% of your time parsing verbose output. By adding a cheap editor model — like a lightweight GPT-4o mini or a local Llama — you can automate that cleanup. Concise output mode only helps a little bit. Tools like this still have a reason to exist. The twist? The very thing that makes AI seem ‘smart’ — its ability to waffle, to consider, to hedge — is actually a bug. We’ve been trained to think that more words mean more intelligence. But in practice, conciseness is a superpower. The future of AI interaction is not about making the AI talk more like a human. It’s about making it talk less. And to do that, you need a translator.
So stop treating AI like a friend. Treat it like a worker. Hire a second AI to be its editor. Stop treating AI like a friend. Treat it like a worker. Hire a second AI to be its editor. The token vomit stops here.
FAQ
Q: Isn't this just adding complexity and cost? Why not just use a better prompt?
A: Better prompts help, but they don't solve the fundamental issue: the model's training data biases it towards verbosity. A separate editor model is cheaper than you think — a GPT-4o mini call costs fractions of a cent. The complexity is trivial to implement with a simple API call. The payoff is huge: you reclaim cognitive load.
Q: What's the practical takeaway for someone using AI daily?
A: If you use AI for coding or writing, try this: generate your output with a powerful model, then pipe it through a lightweight model with a strict 'edit for conciseness' prompt. You'll see a dramatic reduction in fluff. Tools like the 'vomit' repo or 'claudish-to-english' are ready to use. Or build your own with a simple system prompt.
Q: Isn't the verbosity a feature, not a bug? Doesn't it make AI more trustworthy?
A: Verbosity is a crutch. It makes the AI seem thoughtful, but it's actually a sign of poor alignment. The most trustworthy AI is one that gives you the answer you need in the fewest words. The 'editor' pattern is a temporary fix. The long-term solution is better base models. But until then, this is the smartest workaround.