Turn a confident source into separate claims
Bullshit Detector is a set of agent skills for checking a video, article, tweet, PDF, or local draft. It extracts individual claims, searches for independent evidence, and produces a linked verdict for each one.
This matters because an ordinary summary preserves the source’s confidence. A claim report separates facts that can be checked from opinions, missing evidence, and conclusions that need human judgment.
The result includes confirmed, plausible, misleading, false, and unverifiable labels, plus a score for the whole source. Each verdict must cite evidence instead of relying on the model’s memory. The repository also includes a fetching skill that can turn supported articles, videos, social posts, and PDFs into clean text before the check starts.
Check your own launch draft first
The most practical use is not judging strangers on the internet. Run it against a launch post, sales page, investor update, or README before you publish.
Install the skills with the public skills installer:
npx skills@latest add SerhiiKorniienko/bullshit-detectorThe fetching script also needs uv. Then give your agent the draft and the
primary sources behind it:
Check every factual claim in this launch post. Start with the claims a sceptical customer would challenge. Link the best source for each verdict, and leave private company figures as unverifiable unless I provide evidence.
Fix the claims that are too broad, stale, or weakly sourced. Follow every important citation yourself before publishing. The report is a research pass, not an approval stamp.
What it can fetch
The ingestion skill supports ordinary web articles, YouTube transcripts, TikTok captions, tweets, PDFs, and local files. It uses deterministic scripts to fetch and clean the source, then leaves claim extraction and judgment to the agent.
The repository treats fetched material as untrusted text. Its scripts fence the source and neutralise text that tries to close that boundary. This is an important protection when an agent reads arbitrary pages, but it does not make every downloaded file or linked source safe.
The full setup works in Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and other tools that support portable agent skills and web search. Chat interfaces with restricted web or command access need a paste-driven workflow and cannot use every fetcher.
Do not treat the score as a measurement
The project documents several important limits. It checks premises, not the logic that connects them. A source can contain individually true facts and still reach a bad conclusion.
It can cite only pages that its search and fetch tools can reach. Blocked publishers, copied articles, and search-ranked marketing pages can distort the evidence set. Web results also change, so two runs can reach different verdicts.
The repository does not yet have an evaluation harness that measures how often its verdicts are right. Its published examples are useful demonstrations, but the author selected them and the same system produced the reports. An open issue tracks the missing evaluation work.
These limits do not remove the value of the workflow. They change how you use it. Read the cited source, look for an original document, and keep “unverifiable” when the evidence is not public.
Current pace and alternatives
Bullshit Detector is MIT licensed and has been releasing changes quickly. The recent releases added report consistency checks and moved more bookkeeping from the model into deterministic scripts. Pin a release when a report must be repeatable, because the rubric is still changing.
Use it when a source contains many checkable claims and a manual first pass would take too long. Use a professional fact-checker, subject expert, or legal review when an error could harm someone or create a regulatory risk. The skill can show where to look. A person still owns the decision to publish.