AI is fast and fluent, and fluency is not accuracy. A confident, well-formatted answer can still contain a fabricated case, a wrong deadline, or a "fact" that isn't in the record. Every skill here is built around that reality — and so should your habits be.
Treat every AI output as a lead, never a source. The signature on the filing is yours, not the model's.
Every date, citation, and factual assertion routes back to the underlying record or authority before it leaves your office.
Watch a confident answer collapse under verification, and the verification habit sticks for good.
Be deliberate about which client facts enter any tool. Confidentiality is a decision you make on every prompt.
The skill drafts and flags its assumptions. You verify, you decide, you sign. That step is never delegated.
OPEN and the skills will not assert them. The method, the four substantive defects v0.2.0 fixed, and the eleven errors we made finding them are all published at /verification-protocol. A library that tells you to check its work should show you its own.