Bullshit Detector

Read a real report → — 60 claims from a 2.7M-view "the AI bubble is popping" video, 44 of them individually searched, every verdict linked.

Agent skills that fact-check the internet. Point your agent at a viral YouTube video, article, tweet, or PDF — get a claim-by-claim verification report with sources and a BS score (0–10) instead of taking "10 WAYS TO MAKE MONEY WITH AI 🤯" at face value.

Portable Agent Skills — plain markdown + self-contained Python. They work in Claude Code, Codex, GitHub Copilot, Cursor, Gemini CLI, OpenCode, Zed, and any harness that supports the skills format and has web search. No terminal? Claude Cowork installs the whole bundle from a GitHub URL.

Built in the open with Claude Code — an AI helped build the tool that fact-checks AI hype, and the example report is it auditing its own kind.

Follow @SerhiiFounder for new skills and fact-check experiments, or join the newsletter to get them in your inbox.

Quickstart

1. Install uv if you don't have it (the fetch script uses it to self-resolve its dependencies).

2. Run the skills.sh installer and pick the skills and agents you want:

npx skills@latest add SerhiiKorniienko/bullshit-detector

3. Ask your agent: "is this bullshit? <url>", "fact-check this video", "summarize <url>", "explain the part at 12:30".

Install as a Claude Code plugin

Prefer a managed bundle that updates when a new version ships, instead of copied files you maintain yourself? Inside Claude Code:

/plugin marketplace add SerhiiKorniienko/bullshit-detector
/plugin install bullshit-detector@serhii-korniienko

Two ways to install, two philosophies:

• skills.sh copies the skills into your setup so you can hack on them and make them your own. Works with any agent (Claude Code, Codex, Copilot, Cursor, Gemini CLI, OpenCode, …).
• The plugin keeps them as a read-only, always-current bundle — best when you just want it to work and follow along as it evolves. Claude Code and Claude Cowork.

Pick one, not both — installing both gives Claude Code two copies of every skill.

What works where

Rows are the things you'd ask for, columns are where you're asking. Step-by-step setup per app lives in SETUP.md.

The Claude Code and Cowork columns are field-tested (real runs, latest 14 Aug 2026). The Chat and coding-agent columns are what each platform's docs support — if a cell lies to you, that's a bug, tell me. Coding agents need to be able to run scripts and search the web — per-agent specifics in SETUP.md.

Per-app walkthroughs: Claude Code · Cowork · Chat · Codex · ChatGPT · Copilot · Cursor · Gemini CLI · everything else

Why These Skills Exist

#1: Viral ≠ true

A finance guy with 1M views tells you the "only 14 ways to make money with AI". How much of it is real? Views, production value, and confidence are not evidence. The fix is boring: extract every claim, check each against independent sources, and score what survives. That's exactly the work agents with web search are good at and humans never bother doing.

The fix: bullshit-detector — per-claim verdicts (✅ confirmed / 🟡 plausible / 🟠 misleading / ❌ false / ❓ unverifiable), a hype-signal scan, an incentive analysis ("who benefits if you believe this"), and a 0–10 BS score. Verdicts require sources — the skill forbids confirming or refuting from model memory alone.

#2: Agents can't watch videos

Your agent can't sit through a 27-minute video, and YouTube's official API won't give you captions for videos you don't own. Same story with tweets, where the official API now bills per post, and with paywalled articles.

The fix: fetch-content — one script that turns any URL into clean text + metadata with no API keys: YouTube transcripts and TikTok captions via yt-dlp, articles via readability extraction, PDFs, tweets via free endpoints. Every failure mode produces an actionable hint (paywall → paste, no captions → Whisper) instead of a silent guess.

#3: Separation of fetching and judging

Ingestion and analysis are different jobs. Scripts do the deterministic work (fetch, parse, normalize); the agent does the reasoning (extract claims, search, judge). Because analysis skills only ever see normalized text + metadata, adding TikTok support one day touches zero analysis logic — and the same detector works on a tweet and a 3-hour podcast.

Example

A real run against a 1.16M-view "make money with AI" video: examples/0.4.x/report-14-ways-to-make-money-with-ai.md.

BS score: 5/10 — real tools, real trends, guru math, and a funnel every four minutes. 12 claims verified: 4 confirmed, 2 plausible, 3 misleading, 0 false, 3 unverifiable. Among the catches: "Renaissance, D.E. Shaw, Two Sigma only trade employees' money" (true for one fund of one firm), and marketplace stats sourced from the marketplace's own PR.

And a TikTok run — a 552K-view "our Sun has a hidden twin" video: examples/0.4.x/report-second-sun-binary-star.md (BS score: 9/10 — real astronomy vocabulary stitched onto a fabricated cosmology).

A 137K-view "$1M YouTube channel in 1 hour a day" video — examples/0.5.0/report-1m-youtube-channel.md (BS score: 7/10). The advice is fine and unremarkable; the headline "$76,000 per video" turns out to be total business revenue divided by videos published. Every proof point is a number only the seller can see, which the report says plainly rather than pretending to have audited it.

And the awkward one: a 43K-view video arguing the AI buildout is about to collapse, checked by a tool built with Claude — examples/0.5.0/report-claude-situation-shitshow.md (BS score: 5/10). The reporting holds up; the arithmetic behind its headline number is roughly double reality. The claim it rates ❌ false is also the one most favourable to Anthropic, so the report carries a conflict-of-interest disclosure and links every source to check it against.

Someone on Hacker News asked for the obvious test — run it on this README. examples/0.4.x/report-own-readme.md (BS score: 3/10). It caught a two-year-stale API price and a "30-second setup" that began with installing a package manager, both fixed in v0.4.2, and one thing that can't be fixed by editing: the only evidence this tool is accurate is reports it wrote about videos its author picked.

Check your own draft before you publish

The detector runs on any text, including yours. Point it at a post, README, or launch announcement you're about to ship — "fact-check my draft" — and it flags the claims a hostile reader would go after first, with the source that fixes each one. Cheaper than a correction.

That's how examples/0.4.x/report-own-readme.md exists: someone on Hacker News asked for it live, and it found a two-year-stale API price before more people did.

What it doesn't do

Honest limits, because a tool like this earns nothing by overselling itself:

• It checks premises, not reasoning. Eve