Paper into records

There is a drawer, a folder of receipts, or a tray of odds and ends accumulated over years, and you do not know what is in it. You buy a fourth glue stick because you forgot the three you had. A box is one directory of your data, a card is one markdown file in it, and the agent is the coding agent that turns what you photographed into those cards.

What you do. Open the capture page on a phone, photograph things, and talk while you do it, in as many takes as you like: what a thing is, where it lives, that the pan is the good one for baking. Or drop in a folder of files, or run a stack of pages through a scanner.

What the box does. A capture becomes a capture-session card: a timeline of your transcribed speech with each photo placed where it was taken. A bulk drop becomes an upload-batch card with a manifest. A scanned stack becomes one image card per page, each carrying the text read off it and a description. The agent then extracts one record card per thing: a name, where it is, dates, and links back to the photo and to the words it came from. Asides that fit no field stay on the card as your words.

What it needs. A phone or a scanner, and a box the device can reach. Phone capture, triage, the scanner contract, what it requires.

Where it is still rough. A recorded walk of this exact task (one workroom tray, 2026-08-24) produced twenty reviewed records and answered "do I have a glue stick?" with "yes, four". In that walk the agent wrote counts as prose instead of using the record card's quantity field in eighteen of the twenty. There is no item or inventory card type: an inventory is a set of record cards, and the type named inventory is the built-in Storage screen. The same walk found no screen showing every container and its counts, and implementation words showing through the interface. Capture is owner-gated, and photos rotated a quarter turn render upside down.

What makes it possible

Read next. Triage, capture-session, image, upload-batch, record, pdf.