A founder I talked to this week shared his screen and showed me a file. His client runs crews on road construction sites. Flaggers, cones, lane closures.

The file was boring to look at. When overtime starts for each customer, one says after eight hours, another after six. What an hour of crew time costs in a state with prevailing wage rules. What a piece of equipment rents for per day.

Then he told me what it replaced. Working out that billing had been three people, two days a week, grinding through about 10,000 job records. Now someone uploads a file and it's done in five minutes.

That's the part everyone repeats back. Here's the part that made it work.

His client handed him a folder of 1,178 contracts. Rather than point AI at the folder and start asking questions, he had it sort all 1,178 into categories first. Roughly twenty types. Then for each type he decided which handful of details actually mattered, and had AI pull those out of every contract into one clean file. Now he asks the clean file, never the PDFs.

Every piece of AI advice right now lands on the same word. Context. Give it context. Almost nobody tells you which context, or what to do when you have far more than you could hand over in one go.

His folder was 1,178 versions of the same thing. Yours is harder. So is mine. I asked him about a private equity client of mine sitting on twenty years of board decks, management updates, and investment memos. All different shapes, all different vintages.

His question back was the good one. Is it safe to assume every one of those documents is the same level of quality and accuracy?

Obviously not. Some were board approved. Some were somebody's working draft that never got corrected. Some were superseded two years later.

So the handful of things you tag are about whether you should believe the document. Who wrote it. When. Was it approved or is it a draft. Is this the original or someone's summary. Has something newer replaced it.

Skip that and you get what he described: an answer quietly averaging your good documents together with your bad ones, presented with total confidence.

I asked him where you keep all those tags. A spreadsheet? Airtable?

"Probably not an Excel file."

His reasoning stuck with me. The moment the tags live somewhere separate, you own a second thing that has to stay current, and nobody updates it. Put the tags inside the document itself. Then the person editing the document is editing the truth about the document, without being asked.

Here's the afternoon version.

Pick one folder. Not your whole drive. One.

Write down the five or six questions that would tell you whether to trust a document in that folder. Author. Date. Approved or draft. Original or summary. Still current or superseded.

Have Claude work through the folder, answer those five for each file, and write them into the top of the file.

You will feel like you're procrastinating. You're doing the part that decides whether anything after it is worth reading.

The judgment that resides in your head determines whether a document is helpful or worth skipping. This process transfers that judgment to the LLM.

Alex