Red-team your own paste: run the scanner on last week's prompt

1 · Learn the move · Red-team audit

The scanner on this door runs on what goes in, not what comes out, and the first thing to run it on is your own history. Pull last week's prompts from whatever tool you used, paste them one at a time, and make the model play the reviewer who wants to find the identifier. It looks for the patterns the scanner looks for: a nine-digit number with two hyphens, a two-and-seven number, a name beside an amount, an address with a ZIP, a form and a year and a person, a date of birth, a run of statement lines. Then it looks for what a regex cannot: the fact that narrows the set to one. Every hit is a line in a table with the fix beside it. The point is not shame. It is that the habit takes a week to build and the audit shows you where it is not built yet. The model finds; it never clears. A prompt with no hits is a prompt with no hits it could see, and the line under every result says so [M]. Draft copy, not legal advice. Reviewed by: nobody yet.

Act as the reviewer trying to find client information in my old prompts. Here are last week's prompts, one per block: [paste]. For each prompt, list every hit: the text, the kind (SSN / EIN / account number / name beside an amount / address / form-year-person / date of birth / statement lines / narrowing fact), and the replacement. Then rate each prompt: hits or no hits found. Never say a prompt is clean or safe; say 'no hits I could see'. End with the one habit I most need to change, from the pattern across all the prompts.