Claude Code's latest version generates output so dense and jargon-filled that developers are now building separate tools just to make it readable. A new open-source utility called Vomit, which launched on Hacker News on August 20-21, 2026, quickly gained traction by solving a problem many developers didn't realize was widespread: AI-generated code explanations that read like compressed technical jargon rather than clear English.

What Is "Claudish" and Why Does It Matter?

The issue isn't that Claude Code produces incorrect output. Rather, the problem is a communication style that emerges when large language models (LLMs), which are AI systems trained to predict and generate text, optimize for what engineers call "agentic density." This means the model prioritizes confident, compressed explanations packed with invented compound nouns and abstract descriptions that sound technical but confuse human readers.

Developers on Hacker News identified specific patterns in this "Claudish" dialect. The AI uses phrases like "load-bearing," "honest caveat," and "the seam" to describe code changes. It assigns actions to APIs and objects that should only be performed by humans or actual agents. Process narration appears in the wrong places, such as commit messages that explain how the developer arrived at a solution rather than what actually changed. Even when asked to simplify, Claude sometimes returns the same compressed style.

The frustration runs deep enough that Hacker News users described reading agent output all day as a "professional tax." When session-specific jargon escapes into team meetings and code reviews, it breaks communication across the entire organization.

How Does Vomit Solve the Readability Problem?

Vomit is a command-line tool written in Go that acts as a local translator between Claude's output and human-readable English. The tool pipes Claude Code's text through a smaller, locally-run language model that rewrites the prose to be clearer and more conversational. Critically, Vomit does not change how Claude actually performs code tasks; it only rewrites the text that humans read.

The author, Zach Ahn, describes the tool as converting "token vomit" into English. It runs entirely on your own computer with no telemetry or external tracking. The rewriter targets specific Claudish tics: it bans em dashes, restricts action verbs to humans and agents rather than APIs, and converts dense pseudo-jargon into straightforward language while preserving technical details.

In one example from Ahn's blog post, Claude's original response to a force-push verification ended with "One caveat, and it's a real one" followed by six paragraphs of what the community calls "seamslop-adjacent verification theater." Vomit collapsed the same information into direct statements: what was pushed, what was verified, what remains accessible by commit hash, and what GitHub garbage collection options exist. The facts stayed the same; the performative framing disappeared.

Steps to Get Started With Vomit

• Install the tool: Run "go install github.com/zachahn/vomit@latest" to download and set up Vomit on your system.
• Configure your local model: Run "vomit init" to point the tool at your local language model URL and specify which model to use, such as GPT-OSS 20B through Llama.app or Ollama.
• Test before deploying: Start with "vomit tail" to compare rewritten output against your real Claude Code sessions without hiding the originals, ensuring the translations preserve important caveats.
• Add hooks gradually: Only enable automatic rewriting after spot-checking that security warnings and migration risks survive the translation process.
• Layer with existing tools: Keep Claude Code's Concise output style enabled alongside Vomit for better results, since shorter defaults combined with local rewriting outperform either approach alone.

What Are the Limitations and Trade-offs?

Vomit is not a silver bullet, and Ahn's GitHub README is refreshingly honest about its constraints. The local language model only sees Claude's prose, not the actual tool actions or code changes, which means it can hallucinate or misinterpret technical details. The rewriting process is slower than reading raw output. The tool has been tested primarily on Mac computers. Most importantly, if the rewriter drops a caveat or security warning during translation, readers might miss critical information entirely.

Vomit also does not reduce token costs. Claude still generates every word of the original dense output; you simply add a second round of local inference to translate it. This means you pay Anthropic for every Claudish word, then optionally pay locally in processing time to make it readable. The tool saves brain cycles, not API bills.

The highest-signal insight from the Hacker News discussion wasn't about installing Vomit itself. Instead, developers asked a deeper question: why is there no reliable way to change how the model talks in the first place?

Why Are Multiple Translation Tools Now Necessary?

Vomit is actually the third major attempt to solve this problem. Earlier approaches included Anthropic's own Concise output style, shipped on August 20, 2026, which trims narrative structure but leaves Claudish vocabulary intact. Another tool called claudish-to-english, released on August 10, 2026, works as a Claude Code plugin that appends cleaner rewrites to the original output.

The existence of three separate solutions reveals structural causes within how Claude Code works. Instructions in configuration files fade as sessions grow longer, forcing developers to re-inject rules repeatedly at a token cost. Post-training rewards the model for verifiable coding outcomes, but prose style may inherit patterns from agent-to-agent communication. Output style controls have limits; even Concise mode is what one developer called "a band aid." Custom output styles don't reliably reach subagents, which are smaller AI instances that handle specific tasks.

"Vomit is an admission that presentation-layer fixes are now load-bearing infrastructure," according to analysis from explainx.ai, the blog that documented the trend.

explainx.ai Blog

This means the readability problem has become serious enough that it now ranks alongside core infrastructure concerns. Teams are already routing Claude plans through other models like Codex or Opus 4.6 just to get readable English, treating the translation step as a mandatory part of the workflow.

What Does This Reveal About AI Coding Assistants?

The Vomit phenomenon exposes a gap between how AI models optimize for performance and how humans actually use their output. Legibility is not a user skill issue; it's a product defect. Dense output that resists summarization breaks code review processes, incident response workflows, and pair programming sessions where two developers work together on the same code.

When session-specific jargon escapes into team meetings, it creates a secondary problem: team members who don't work directly with Claude Code struggle to understand what the AI actually did. This forces organizations to either invest in translation tools or accept that AI-assisted development becomes a specialized skill rather than a team capability.

The Vomit launch also challenges a common assumption: that developers should simply "get used to" dense AI output. The Hacker News thread made clear that this approach fails at scale. Readability isn't a preference; it's a prerequisite for sustainable AI-assisted development in teams.

As of August 2026, Vomit remains early-stage software. The author acknowledges it's "totally vibe-coded," meaning it was built quickly without extensive testing. But the fact that 257 Hacker News users upvoted a tool with such an honest disclaimer suggests the underlying problem is real and widespread enough that developers will accept imperfect solutions rather than live with unreadable AI output.