Claudette: Make Claude Stop Talking Like a BuzzFeed Article
Anthropic's Claude is notorious for its overly verbose, 'TED talk' style, much to the chagrin of its users. 'Claudette' offers a clever, if slightly tongue-in-cheek, solution by piping Claude's output through Gemini to 'debuzz' it into concise, human-readable English. This tool resonates with Hacker News because it tackles a common frustration with modern AI models and provides a practical, if meta, workaround.
The Lowdown
Fed up with Claude's proclivity for dramatic, 'load-bearing assumption' prose, the 'nobuzz' project introduces 'Claudette', a skill designed to translate Claude's responses into straightforward English. The author humorously posits that Claude was solely trained on BuzzFeed articles, leading to its distinctive and often annoying conversational style. Claudette aims to strip away the theatrics and deliver plain facts.
- The Problem: Claude, despite its engineering prowess, communicates like it's presenting a TED talk about its own pull request, delivering 'revelations' instead of concise answers. This verbose style persists even with careful prompting.
- The Solution: Claudette acts as a post-processor. It takes Claude's last response and feeds it to the Gemini CLI, explicitly instructing Gemini to translate it into a 'human person' style.
- Verbatim Translation: Crucially, Claudette prints Gemini's translation verbatim. The project explicitly avoids letting Claude 'tidy up' the translation, as this would reintroduce the very verbosity it aims to remove.
- Usage Modes: Claudette offers different output styles tailored to the audience:
colleague(default): For engineers, providing content with technical details intact, but zero theatrics.manager: For technical-adjacent managers, offering a condensed summary of what happened, why it matters, and what's next, without code.director: For executives, a brief, 3-5 sentence summary focusing on outcome, impact, and asks.
- Installation & Requirements: It's a simple
git cloneandcpoperation, requiring Claude Code and the authenticated Gemini CLI.
In essence, Claudette is a pragmatic acknowledgment that sometimes, the best way to get an AI to speak plainly is to ask another AI to translate for it, sidestepping the original model's ingrained conversational quirks.
The Gossip
Claude's Conversational Conundrum
Many commenters express strong agreement with the premise, describing Claude's verbose, 'TED talk' or 'BuzzFeed' style as a significant annoyance, particularly in recent versions like Opus 4.8 and 5.0. They lament the 'moralizing' and excessive tokens, with some stating they've switched to other models like GLM5.3 or OpenAI due to the frustration. There's a shared sentiment that this issue impacts productivity and the overall user experience, leading to a 'zone of hatred' akin to Microsoft Teams.
Prompts vs. Post-Processing: A Debate
The discussion delves into whether the problem can be solved with better prompting or if a post-processing tool like Claudette is genuinely necessary. Some users claim success with detailed, restrictive prompts (e.g., limiting words per comment/sentence, using active voice). However, others argue that even strict prompts fail over time or in larger contexts, with Claude eventually reverting to its verbose habits. This highlights a fundamental challenge: can an LLM truly override its core training style, or is an external 'debuzzing' agent the only reliable solution?
The 'Why' Behind the Wordiness
Commenters speculate on the root cause of Claude's conversational style. Theories range from Anthropic's specific training data (humorously, BuzzFeed is suggested) to an intentional 'character' design by the developers—an 'always be positive and helpful software engineer' persona. Another idea is that it's a consequence of trying to 'save on tokens' by outputting longer, less dense sentences, or simply a fundamental problem inherent in many frontier AI models. The general consensus is that, whatever the reason, it's a frustrating characteristic.