verify-claims

A document can read as confident and still be full of guesses. The expensive failures are never typos—they are confident claims that trace to nothing: a regulation number copied from a secondary news article, a total nobody recounted, a URL that was never actually fetched, a fact that survived from an earlier draft. One fabricated detail costs more trust than the whole document ever earned.

verify-claims follows a five-step method: classify every checkable claim, rank them by how much a reader would act on them, trace each to its strongest available source, verify what is verifiable in the current run—fetching links for real, recounting totals from source data, opening the cited section—and honestly label the rest as unverified. The principles are straightforward: primary sources beat secondary, totals must be recounted not transcribed, and a link that was never fetched does not exist.

It ships as a single SKILL.md file with zero external dependencies, disciplining whatever tools the agent already has. It is well suited for legal and compliance documents, research reports, technical specifications, and financial analyses where sources matter and readers make decisions based on the content. It is honest about its limits: verified means traced to a source this run, not proven true; paywalled sources are labeled unverified (walled); recounting validates arithmetic and provenance, not the underlying data collection.