916 products, 30 agents, one operator

The distributor's 916 WooCommerce products had drifted from a separate 250-page product guide. The export was awkward: six columns, specs buried in HTML, and a hashed SKU field with the actual part number in the title. A preliminary review of 554 guide entries had found 133 discrepancies and 71 links to a retired URL slug. The guide had errors too, so it couldn't automatically win every disagreement. The person running the audit wasn't an engineer.

I packaged the audit as a master prompt the operator could paste and run. Extraction agents read the guide in page batches and visually recheck sampled pages. A verification pass checks each field against its source. Reviewer agents then try to disprove each flag by rereading the cited page. They can't infer specs from filenames or fill gaps by guessing. Suspected errors in the guide go to the client for a decision. The process allows about 30 agents and pauses for three screenshare checkpoints with me.

The operator completed the comprehensive run herself. The process produces an extended workbook, a proposed-fixes CSV keyed by product ID and an executive summary. The master prompt includes the phased plan and baseline flag categories, with a fallback if orchestration is unavailable. She runs it on her own data; I join at the three decision points.