Glonce

New Claude Code skill runs Claude's replies through Gemini to cut buzzwords

toolsThe Glonce editorsHacker News🔥67Glonce rating67Impact35Novelty60Relevance85Surprise70Players55Credibility85
New Claude Code skill runs Claude's replies through Gemini to cut buzzwords

A developer who goes by aakil on Hacker News published a Claude Code skill that rewrites Claude's own replies to sound less like marketing copy. The project, hosted at github.com/adnanakil/nobuzz and invoked with the command /debuzz, takes Claude's last response and hands it to a different model, Gemini, through Google's Antigravity CLI (agy), whose only job is to restate the text the way a person would. The README jokingly proposes calling the project "Claudette" but says it is explicitly not called that, citing a person named Cat Wu, described only as "famously litigious," as the reason; the source gives no further detail on who she is. The README opens with a joking claim that Claude was trained solely on old BuzzFeed articles, offered as a tongue-in-cheek explanation for the style being fixed, not as a factual claim. The stated target is Claude's tendency to narrate ordinary engineering answers as if delivering a TED talk: numbered "revelations," a solemn note that the third point is the most instructive, and a dramatic "kicker" line. The README gives a before/after example. Before, Claude explains a bug as: "Here's where it gets interesting: the retry logic isn't just a nice-to-have, it's the load-bearing assumption of the entire sync pipeline," building to three numbered points and a final line calling the last one "the kicker." After running through /debuzz in "colleague mode," the same explanation becomes a flat list: the sync pipeline's retry logic has three bugs, syncQueue.ts:142 swallows an ETIMEDOUT error instead of re-queuing the job, the backoff cap is 2 seconds which is too low for mobile networks, and the dedupe key includes a timestamp so retries are never actually deduplicated, followed by a concrete fix: strip the timestamp from the key, raise the cap to 30 seconds, and re-throw the timeout error. Mechanically, the skill writes Claude's previous reply to a temp file and runs agy -p with the file's contents plus plain-English style instructions, because agy's headless mode does not read stdin and will not read files outside the project directory; it then prints Antigravity's output verbatim. The README states this is deliberate: letting Claude "tidy up" the Gemini translation would reintroduce the exact voice the skill is meant to remove. If agy itself errors, which the README says is usually an authentication problem, the user sees that actual error; Claude only offers its own rewrite as a clearly labeled fallback, avoiding a setup where the model being fixed silently rewrites its own output. Installation requires Claude Code and the Antigravity CLI, installed with a shell script on macOS and Linux or a PowerShell command on Windows, followed by running agy once to complete Google Sign-In; the skill itself is installed by cloning the repository and copying its debuzz folder into ~/.claude/skills. Usage is /debuzz [mode] [text]: with no text argument it translates Claude's previous reply, pasting text after the mode translates that text instead, and it also triggers on natural phrases such as "say that in normal english." The project is released under the MIT license. On Hacker News the submission had reached 241 points and 170 comments at the time it was captured.

Key facts

  • The skill runs as /debuzz inside Claude Code and is only informally nicknamed "Claudette"; the README explicitly says that is not its real name.
  • It sends Claude's last reply to Google's Gemini model via the Antigravity CLI (agy) and prints Gemini's rewrite verbatim, without letting Claude edit it.
  • The README's before/after example turns a dramatized bug explanation with a numbered "kicker" into a flat three-bug list plus a concrete fix.
  • If the Antigravity CLI errors, usually on authentication, the user sees the raw error; Claude offers its own rewrite only as a clearly labeled fallback.
  • Setup requires installing the Antigravity CLI and signing in with a Google account, then copying the skill's debuzz folder into ~/.claude/skills; the project is MIT licensed.

Why it matters

Claude Code users on Hacker News have long complained that Claude explains routine engineering answers with an inflated, narrative style, numbered "revelations" and a dramatic closing "kicker" even for a plain bug fix. This project treats that as a problem prompting alone cannot fix and routes the output through an entirely separate model instead. Rather than trying to prompt Claude out of the habit, it accepts the habit as persistent and outsources the final rewrite to Gemini, then commits to printing that rewrite unedited so Claude cannot reintroduce its own voice on the way back out.

Who it affects

Claude Code users who want plain, direct technical answers, especially for routine debugging explanations, are the direct audience. It only helps someone who already has both Claude Code and a working Antigravity CLI account, so it does not reach Claude users outside that setup.

How to use it

Install the Antigravity CLI (curl script on macOS and Linux, a PowerShell command on Windows) and run it once to complete Google Sign-In, then clone the adnanakil/nobuzz repository and copy its debuzz folder into ~/.claude/skills. The command is /debuzz [mode] [text]: with no argument it rewrites Claude's previous reply, pasted text after the mode is rewritten instead, and natural phrases like "say that in normal english" also trigger it. It is free and released under the MIT license; the only cost is whatever the Antigravity CLI and Gemini access require on Google's side, which the README does not specify.

How solid is it

This is a single-developer, open source project posted directly to Hacker News, not a first-party Anthropic or Google tool, and the README carries no adoption numbers, download counts, or outside testimonials. Its evidence for effectiveness is the one before/after example the author supplies. At the time this story was captured, the submission had 241 points and 170 comments on Hacker News, indicating attention on the site but not independent verification of how well the technique works.

Risks and caveats

The tool depends on a second CLI and a separate Google account sign-in, adding setup steps and an external dependency to every Claude Code session that uses it. The author's own design note flags the core risk directly: letting Claude "tidy up" Gemini's rewrite would quietly reintroduce the voice being removed, which is why the skill prints that output verbatim and surfaces raw Antigravity errors instead of silently falling back to Claude rewriting itself. The opening line about Claude being trained "solely" on old BuzzFeed articles is a joke in the README, not a factual claim about how Claude was trained.

“it hands the response to a different model, Gemini (remember gemini from last summer?) via Google's Antigravity CLI, whose only job is to say it like a human person”

— project README, adnanakil/nobuzz