Not negotiable. No prompt here may generate a testimonial, review, case-study quote or customer story. The FTC rule at 16 CFR 465 never mentions AI and does not need to: it prohibits misrepresenting that a reviewer exists or that they used the service, which is what a generated testimonial does by construction. Helping a real customer tidy a review of something they actually used is a different thing and is fine.
Three more. No statistic or dollar figure ships without a human-verified source; run the claim substantiation extractor on every asset before it reaches a customer. Label synthesised voice or likeness and get consent, noting that consent does not remove the labelling duty. And marketing materials we hand a client are our exposure too, under the means-and-instrumentalities doctrine that reaches agencies.
Changing a prompt. Re-run it on three to five real briefs where we already know the right answer. Grade pass or fail plus a written note on why, never a one-to-five score. Teams collapse scales to binary anyway, and the written critique is where the nuance lives.
The same 82 are in the vault folder as individual files, with a project instructions file for loading the whole thing as project knowledge.