Make It Find What It Missed

The most dangerous thing about AI output is how finished it looks. First drafts arrive formatted, confident, and wrong in places, and nothing about the formatting tells you where. The difference between people who get burned by that and people who don’t isn’t smarter prompting on the way in. It’s pressure on the way out.

Two stories from this month’s Roundtable, then the protocol.

A member in a regulated field took a mandated 36-page assessment, the kind of form that makes grown professionals sigh, and had AI condense its score sheet to eight pages. Then he did the thing most people skip: he asked it, twice, whether the condensed version thoroughly covered everything the original required. On the second pass, the AI caught its own error: it had quietly merged three required domains into one because that seemed tidier. It confessed and fixed it. The condensed form was real; the first version of it wasn’t done.

And the one I use on every complicated deliverable: I find one error myself, then I say, I found an error and I’m not going to tell you what it is. Your analysis wasn’t thorough enough. Go find all the errors. The quality of what comes back after that sentence is a different tier, because now it’s actually hunting.

The four-ways protocol

For anything where accuracy genuinely matters, one check isn’t a check; it’s the same blind spot twice. So I require four different METHODS, and I name them: read the source with your text extraction, then verify against the original using your vision on the actual images, then do a logic pass, does each output line make sense against the rules we know, then cross-check totals and patterns against expectations. Four different instruments pointed at the same work find different mistakes, and something better happens across rounds: the AI starts learning where it makes its mistakes, and the error count drops toward zero instead of hiding.

The finishing move for anything at scale: demand a confidence log. Every line the AI wasn’t sure about goes on a list for human eyes. You don’t review a thousand rows; you review the eleven it flagged, which is exactly how this protocol turned a member’s document pile into a trustworthy tool (that story’s here: https://tablelandpartners.com/build-the-tool-not-the-task/).

The habit under the protocol

Strip away the mechanics and it’s one habit: never accept the first confident answer on anything that matters. Ask: “does this cover everything”, ask it again differently, make it re-derive its result another way, and reward it for finding its own mistakes rather than for sounding sure. If that sounds like managing a talented new employee, that’s exactly what it is, and it pairs with the five rescue prompts I keep for everyday answers: https://tablelandpartners.com/five-prompts-better-ai-answers/. .