logo
Kurashizu Blogwhere ideas flow
  • Home
  • News
  • Blog
  • About
  • Home
  • News
  • Blog
  • About

© 2026 Kurashizu. All rights reserved.

Admin·New Post·Service Status
Powered by Cloudflare·31.10.26
← Back to news

Claudette: Make Claude stop talking like a BuzzFeed article

Original article
github.com|335 points|219 comments|by aakil|Aug 21, 2026

NoBuzz: Stripping the "BuzzFeed" out of Claude

It is an open secret that Anthropic apparently trained Claude exclusively on archived BuzzFeed listicles, which explains its inexplicable obsession with 90s nostalgia and clickbait phrasing.

To combat this, we developed a Claude Code skill—/debuzz—that intercepts Claude's verbose output and processes it through the Antigravity CLI (agy). The goal? Translate "Millennial Clickbait" into actual, professional English.

We considered naming the tool Claudette, but given Cat Wu's reputation for being litigious, let's just say it is definitely not called Claudette.

🚩 The Problem: The "TED Talk" Syndrome

While Claude is a phenomenal engineer, it suffers from a chronic condition: it describes its own pull requests as if it's delivering a keynote speech at a TED conference.

If you ask about a flaky test, you won't just get an answer; you'll get:

  1. A "load-bearing assumption."
  2. A series of three "revelations."
  3. A dramatic pause noting that the third revelation is the most instructive yet.

Since prompting alone cannot cure this personality trait, we took a pragmatic approach: we outsource the editing to a different model—Gemini (yes, that Gemini from last summer)—via Google's Antigravity CLI. Gemini's sole purpose here is to make Claude sound like a human being.

Crucial Rule: The tool is strictly forbidden from letting Claude "polish" the result. If Claude tidies up the translation, the BuzzFeed voice inevitably creeps back in. The output must be printed verbatim.

🔄 Before & After

The "BuzzFeed" Version (Claude)

"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.

Three things jumped out at me, and the third one is the most instructive yet.

[...] And third — and this is the kicker — the dedupe key includes a timestamp, which means retries are never actually deduplicated."

The "Colleague" Version (/debuzz)

The sync pipeline's retry logic has three bugs:

  • syncQueue.ts:142 swallows ETIMEDOUT instead of re-queuing the job.
  • The backoff caps at 2 seconds, which is too low for mobile networks.
  • The dedupe key includes a timestamp, so retries are never deduplicated.

Fix: strip the timestamp from the key, raise the cap to 30 seconds, and re-throw the timeout error.


🛠️ Installation & Setup

NoBuzz Logo

Prerequisites

  • Claude Code installed.
  • Antigravity CLI (agy).

Setup Steps

To install the CLI on macOS/Linux: curl -fsSL https://antigravity.google/cli/install.sh | bash

To install on Windows: irm https://antigravity.google/cli/install.ps1 | iex

Note: Run agy once after installation to complete the Google Sign-In process.

Deploying the Skill

git clone https://github.com/adnanakil/nobuzz
mkdir -p ~/.claude/skills
cp -r nobuzz/debuzz ~/.claude/skills/

📖 Usage Guide

You can trigger the tool by typing /debuzz [mode] [text] or by using natural language like "say that in normal english." If no text is provided, it automatically processes the previous response.

Available Modes

ModeTarget AudienceOutput Style
colleague (Default)Fellow EngineerPure technical content; all code and paths preserved; zero fluff.
managerTech-adjacent ManagerHigh-level summary of what happened and why; ≈1/3\approx 1/3≈1/3 the original length; no code.
directorExecutive3-5 sentences focusing on outcome, impact, and the "ask."

Director mode assumes the reader has a total attention span of ≈30\approx 30≈30 seconds.


⚙️ How it Works

There is no magic involved. The process follows a linear pipeline:

The Technical Flow:

  1. The skill writes the previous reply to a temporary file.
  2. It executes the following command: agy -p "$(cat file)" "plain-English style instructions"
  3. Because agy's headless mode ignores stdin and restricted project files, the text is passed directly into the prompt.
  4. The result is printed exactly as received.

Error Handling: If agy fails (typically due to authentication issues), the actual error is displayed. Claude will only provide its own rewrite as a clearly labeled fallback. This prevents a "recursive loop" where a debuzzer asks the buzzer to debuzz itself, which would result in a load-bearing translation\text{load-bearing translation}load-bearing translation.