Why should I audit my documents before AI builds from them?

I asked an AI to turn a book I wrote in February into a web app. Before it wrote a line of the build prompt, it read my own folder back to me. The book said "right now, in early 2026" in four places. Two of my documents priced the same offer differently. One chapter leaned on a testimonial from a person I never named.

Seven months of drift, sitting in files I thought were finished.

What goes wrong when AI builds from old files?

A build prompt is only as current as the oldest document it was written from. Hand an AI a stale source and it will build something polished on top of every stale claim, because it has no way to know the claim aged. The fix is a step I skipped for years, in a hurry to see the shiny thing. Before you ask for the build, ask for the review.

The review, step by step

The inventory. Point the model at the folder and ask for a table with two columns: what each asset is, and what it commits you to. That second column matters. A pricing sheet commits you to prices. An old training deck commits you to a promise about what the client receives. Seeing the commitments side by side is where the contradictions show up.

The flags. Ask for every place the documents disagree with each other, with today's date, or with a fact that can be checked. Each flag gets a confidence number. Mine came back with pricing at 99 percent and the stale dates at 97. Those go to the top of the document, above the build prompt, so they cannot be scrolled past.

The promise-to-source table. Every feature in the build gets a row: the promise, the document it came from with its date, and where it shows up in the finished product. Nine chapters, from the February book. Four offer cards, from the July pricing doc. Seven free prompts, from chapter six. A tagline in the footer, from the brand file. If a feature cannot name its source, it is an invention, and inventions get cut or confirmed before anything goes live.

The open list. Mine held four questions nobody but me could answer. The model stopped there and waited.

Why speed raises the stakes

The build tools got fast enough that a working app takes an afternoon. Speed moved the risk upstream. The expensive mistake used to be a bad build. Today it is a good build of a wrong document, live and public by Friday, with a price on it you retired in July.

The review costs a few minutes and one extra prompt. It also changes who makes the decisions. When two rows of the inventory disagree, you choose which one is true. Skip the review and the machine chooses for you, usually by picking whichever version it read last.

Do this today

1. Pick one product or deck you plan to rebuild this quarter.
2. Open the folder it will be built from.
3. Ask your AI for the two-column inventory: asset, and what it commits you to.
4. Read the second column out loud.
5. Mark every place two rows disagree, and decide which version is true before any build prompt gets written.