Bullshit Detector: Open Agent Skills That Fact-Check Viral Videos

A new open-source project called Bullshit Detector packages fact-checking as portable "agent skills" — plain markdown plus self-contained Python — that run inside Claude Code, Codex, OpenCode, or any harness with web search support. Instead of taking a viral video's claims at face value, the agent extracts every claim from a YouTube video, TikTok, article, tweet, or PDF, checks each against independent sources, and returns a report with per-claim verdicts and a 0-10 BS score.

The toolkit deliberately separates ingestion from analysis. A fetch-content script pulls clean text and metadata from any URL without API keys — YouTube and TikTok transcripts via yt-dlp, readability-based article extraction, PDFs, and tweets — while a Whisper fallback handles caption-less videos. Because analysis skills like bullshit-detector, summarize, and explain only ever see normalized text and metadata, adding support for a new source type never touches the reasoning logic.

For engineers, the project is a concrete pattern for building portable, harness-agnostic agent tooling under the emerging "skills" format, with an explicit rule that verdicts must be backed by sources rather than model memory alone. It installs via a one-line npx command or as a Claude Code plugin, and is released under the MIT license.