← ClaudeAtlas

verification-coveragelisted

Audit a drafted conclusion, report, review, or plan for load-bearing claims that were never actually checked — classify each as verified/inferred/assumed, run the cheap falsifying command for the unverified ones, correct what turns out wrong, report a coverage figure, and for a durable artifact emit the checks as a runnable claim suite beside it. Use before delivering an output someone will act on, or when asked "how do we know this is right?", "how confident are you?", "check your claims", or "did you verify that?".
justinmchoi/ai-toolkit · ★ 1 · AI & Automation · score 78
Install: claude install-skill justinmchoi/ai-toolkit
# Verification Coverage An audit pass over a **drafted, not yet delivered** output. It does not make claims true. It makes unverified claims *visible*, which is the part that is currently invisible: in a finished document, a fact confirmed by running a command and a fact assumed from a code comment look exactly the same. ## Why this targets absence claims specifically Reviewing real sessions, the errors cluster hard. They are not bad reasoning, bad judgment, or bad domain knowledge. They are overwhelmingly **existence and absence claims made from having looked in one place**: | Claim asserted | What would have falsified it | |---|---| | "there are no hooks in the toolkit" | `ls <toolkit>/hooks` | | "that module doesn't exist in the repo" | `git ls-tree -r --name-only origin/main` | | "the installed copy is missing its script" | `readlink -f` | | "this field is handed to the downstream service" | grep the consumer | Each cost one command. None was run before asserting. This is a tractable target because **negative and universal claims are exactly the ones that are mechanically falsifiable**, and the ones most likely to be made from partial evidence — you assert a positive because you just saw the thing, but you assert an absence because you *didn't* see it, which is a much weaker warrant. ## Procedure ### 1. Extract the load-bearing claims A claim is load-bearing if the reader would **do something different** were it false. Recommendations, blockers, "X causes Y", cou