Raemon's 'Direct Quotes' skill redesigns AI outputs for instant verification

Raemon, writing on LessWrong, argues that AI hallucinations won't disappear soon, so we should design AI outputs to be trivially verifiable. Instead of summarizing or guessing, an AI should hunt down primary sources and respond with exact quotes, complete with source names and context. A custom 'Direct Quotes' skill for Claude implements this: the AI truncates quotes for readability but never invents text, and a fast dumb AI checks that quotes are accurately transcribed. The UI lets users expand any quote into the full document for deeper inspection.

For complex inferences that require original explanation, Raemon suggests a two-phase approach: the AI gives a quick memory-based answer, then immediately searches for primary sources and generates a new response backed by verified quotes. Inferences themselves are clearly labeled as distinct from sources. Raises the bar for AI-assisted intellectual work by expecting cross-checking with corroborating or disproving evidence. This shift aims to create an internet where AI claims always trace back to a human author, reducing slop and making fact-checking as frictionless as clicking a quote.