Dev builds a Gemini-powered Claude Code skill to strip out the clickbait voice

A developer has published an open-source Claude Code skill called nobuzz (invoked as /debuzz, jokingly branded “Claudette”) that rewrites Claude’s chattier output into plain, direct English. The complaint behind it is stylistic rather than technical: Claude often frames straightforward answers as a narrative, complete with “load-bearing assumptions,” numbered revelations, and a final “kicker,” when the user just wanted to know why a test was flaky. The author’s premise is that no amount of prompting reliably suppresses this voice, so the tool sidesteps the problem instead of trying to fix it from the inside.

The mechanism is deliberately simple. The skill writes Claude’s previous reply to a temporary file, pipes it through the Gemini CLI with plain-English style instructions, and prints Gemini’s rewrite verbatim. The key design decision is that Claude is not allowed to touch the translation — letting it “tidy up” Gemini’s output would reintroduce exactly the tone being removed. If Gemini fails (typically an auth error), the user sees the real error, and Claude only offers its own rewrite as a clearly labeled fallback. Setup requires Claude Code plus an authenticated Gemini CLI, and usage is flexible: with no argument it debuzzes the last reply, or you can paste text to translate, and it also responds to natural phrases like “say that in normal English.”

The project is MIT-licensed and the writeup is tongue-in-cheek, but it points at a genuine friction in agentic coding tools: model “personality” and verbose framing can get in the way of technical communication. The more interesting pattern here is using one model as an adversarial editor for another’s prose — a lightweight, composable way to enforce a house style that self-editing by the same model tends to undermine.