Urgent.News

What's breaking now, across thousands of outlets.

AI

Claudette: Make Claude stop talking like a BuzzFeed article

Article URL: https://github.com/adnanakil/nobuzz/blob/main/README.md Comments URL: https://news.ycombinator.com/item?id=49388752 Points: 217 # Comments: 154

Anthropic trained Claude exclusively on old BuzzFeed articles, which explains Claude's penchant for 90s nostalgia. To address this, Claude and the reporter created a tool called "Claudette" (/debuzz) that translates Claude's responses from millennial clickbait style to regular English using the Antigravity CLI (agy). They plan to name it "Claudette," but avoid the name due to Cat Wu's litigation history.

Claudette is designed to acknowledge that no amount of prompting can fully cure Claude's tendency to sound like a TED talk presenter. Instead of attempting to fix the issue with each prompt, Claudette hands off the response to a different model, Gemini, which translates Claude's words in a more natural, human-like manner.

The retry logic in Claudette's sync pipeline is critical. It includes three bugs: swallowing ETIMEDOUT instead of re-queuing the job, a backoff cap of 2 seconds that's too low for mobile networks, and including a timestamp in the dedupe key, preventing retries from being deduplicated. The reporter outlines the fixes: stripping the timestamp from the key, raising the cap to 30 seconds, and re-throwing the timeout error.

Claudette operates by writing its previous reply to a temporary file and then running agy -p $(cat file) in plain English style instructions. It uses agy's headless mode, which doesn't read stdin or files outside the project, to directly input the text into the prompt. If agy encounters an error (usually authentication), the actual error is displayed, while Claude only offers its own rewritten version as a clear fallback, ensuring the debuzzer doesn't inadvertently promote the original Buzzfeed-style language.

Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at github.com →

More in AI

How Claude’s AI Watermark Actually Works

When people hear “AI watermark,” they often imagine something visible: a hidden character, a special symbol, or metadata embedded inside every sentence. Text watermarking works differently.

World's largest open library calls for volunteers to scan and preserve physical books as AI companies buy, scan, and destroy them — Anna's Archive says ‘time is running out’ as ‘knowledge is permanently monopolized on private servers’

A volunteer for Anna's Archive is calling for volunteers to scan and upload books to the shadow library to help preserve human knowledge for the public. The move comes as more AI companies buy, scan, and destroy books to feed to AI models, which is easier and faster than scanning the written works in a non-destructive manner.

Making Local AI Tool Calls More Reliable

Making Local AI Tool Calls More Reliable While testing our local-first AI assistant, I found an intermittent problem: the model sometimes answered with plain text instead of calling the tool needed to read or update local data. The original system used tool_choice="auto" .

  • Introduced safety net recovery mechanism for local AI tool calls.
  • Assistant retries with toolchoice= required before any tool execution.
  • Modified recovery flow: auto - required - auto during live streaming.

More from Friday 21 August →