Bullshit Detector

Views, production value and confidence are not evidence. Bullshit Detector is a set of portable Agent Skills that do the boring work nobody does by hand: extract every discrete claim from a piece of content, search each one independently, and report what survives. Point your agent at a YouTube video, TikTok, article, tweet or PDF. You get back a claims table - every claim marked confirmed, plausible, misleading, false or unverifiable, every verdict linked to the source it rests on - plus a hype-signal scan, an incentive analysis ("who benefits if you believe this"), and a 0-10 BS score. The core rule is that verdicts require sources. The skill is forbidden from confirming or refuting a claim from model memory, even when the model is certain. Reprints of one wire story are collapsed into one origin, so "widely reported" gets counted honestly. A script recounts the finished report's own arithmetic and rejects it if the tally doesn't reconcile. It works in Claude Code, Codex, OpenCode and any agent harness with web search - plain markdown and self-contained Python, no API keys, MIT licensed. Seven skills ship together: bullshit-detector, fetch-content, coverage-check, summarize, explain, report-card and share. What it doesn't do, stated up front: it checks premises, not reasoning; it can only cite sources it can reach; and it has no eval harness yet, so there is no accuracy number to quote