Turn a Dropbox folder of PDFs into approved spreadsheet rows

Build me a document review desk. Today my team does this by hand with a folder of PDFs open next to a spreadsheet, retyping the same fields over and over. I want one screen where a reviewer works through a queue of documents, checks the fields that were pulled out of each one, corrects anything wrong, and approves it into our Google Sheet.

The queue view is the home screen. Use the Dropbox List Folder operation on a folder path I configure in settings, and show every document in that folder as a row with its file name, its last modified date, and its review status of needs review, in progress, or approved. Approved documents drop out of the main queue and move to a separate Done tab. Keep the review status, the extracted values, and the approval record in the app's own storage, keyed by the Dropbox file id rather than the path, so renaming or moving a file in Dropbox does not lose its history. Never modify or delete anything in the Dropbox folder itself, since the app only reads from it.

Opening a document from the queue takes the reviewer to a detail view. Fetch the file with the Dropbox Download File operation and send its contents to Claude using the Anthropic Create Message operation, together with my extraction instructions and my list of fields. Pass the downloaded contents directly in the Create Message request. Do not try to upload the document to Anthropic first, because our Anthropic connection has a list files operation but no upload operation, so inline contents are the only path that works.

Ask Claude to return structured output where every field carries three things: the extracted value, a confidence level of high, medium, or low, and a short verbatim snippet of the document text the value came from. Confidence is something the extraction prompt reports per field, not a number read from any API, so make that part of the instructions you send. The default fields should be vendor, document date, total, and payment terms, but the field list is mine to edit.

Lay the detail view out as two panes. On one side show the document text so the reviewer can read the actual evidence. On the other side show every extracted field as an editable input, pre-filled with Claude's value, with its confidence marked and its supporting snippet visible next to it so the reviewer can confirm a number without hunting through the whole document. Low confidence fields should be visually obvious. The reviewer can edit any field freely, including ones marked high confidence.

Sort the queue so anything Claude flagged as low confidence comes to the top. A document with at least one low confidence field outranks a document with none, and among those, more low confidence fields sorts higher. Everything else falls below that, ordered by document date. The reviewer should always be spending their attention on the shakiest extractions first.

Approving is the only thing that ever writes to the spreadsheet. When the reviewer clicks approve, read the destination sheet's header row first with the Google Sheets Get Values operation so the columns line up with my actual headers rather than a hardcoded order, then append the approved values as a new row with the Google Sheets Append Values operation. Include the approver's name and the approval timestamp in the row, and store them against the document in the app as well. Then mark the document approved so it leaves the queue.

Two rules matter more than anything else here. Never export a row that nobody approved: there is no bulk export, no automatic export, and no path to Append Values other than a person clicking approve on a document they have looked at. And never append the same document twice, so if approve is clicked again or the page is reloaded mid-save, check the stored approval record first and do not write a second row.

Give me a settings screen where I can edit the extraction instructions and the list of fields, and persist those so they apply to future extractions. From a document's detail view I also want a rerun button that re-extracts that single document against the current instructions, replacing only that document's pending values. A rerun must never touch other documents in the queue, and must never alter a document that has already been approved.

Cache each document's extraction result once it has been generated, so reopening a document from the queue shows the saved values instead of calling Claude again. Re-extraction should only happen when someone explicitly clicks rerun. The Done tab should list approved documents with their final field values, who approved each one, and when.