Fact Check

The problem

Manually verifying claims in news drafts, reports, or AI-generated content is slow and prone to oversight. Relying on search engine snippets often leads to circular reasoning where the source being checked is the one providing the verification.

What it does

• Extracts every factual claim including dates, amounts, and entities using a "verbatim capture" principle to ensure no detail is missed.
• Forces external cross-referencing against a tiered source hierarchy, prioritizing official government filings (T0) and authoritative media (T1).
• Detects sophisticated misinformation tactics like time-shifting (using old news as current), title-body mismatch, and AI hallucinations.
• Generates structured verification reports with deep-link evidence, categorizing claims as accurate, incorrect, biased, or unverifiable.

Frameworks & tools

Python 3, Cue API, and official financial/regulatory data sources like SSE, HKEX, and EDGAR.

Why this beats prompting it yourself

Standard LLM prompting often results in internal hallucinations or "lazy" verification that accepts surface-level search results. This skill implements a strict anti-circular-reasoning firewall and a forced deep-search protocol that bypasses homepages to find primary source documents.

Use cases

• Fact-checking PR drafts against historical financial reports and official corporate filings.
• Verifying AI-generated industry analysis to catch hallucinated statistics and fabricated market data.
• Vetting social media content for "zombie news" where old events are rebranded as breaking stories.

Known limitations

Does not verify subjective opinions, private data, or content behind paywalls. Not designed for real-time breaking news where official sources haven't yet commented.