Summarise the documents people email you

Contracts, statements, reports and forms arrive as email attachments and go unread, because opening eight PDFs to find out which one matters is nobody's favourite hour. This reads each one as it arrives, writes a short summary into a running Google Doc, and emails you one note at the end of the run listing what came in.

Every document is asked the same eight questions - what kind of document it is, who issued it, who it is addressed to, the date on it, its reference number, what it asks the reader to do, any deadline, any total. The answers come out of the document itself, and a question the document does not answer is written down as 'not stated in the document' rather than filled in. The summary is then written from those answers and nothing else: the model never sees the document, the email, the sender or the subject line, so it has nothing to embroider with. It is a summary of what the document says, not an opinion about it, and it offers no advice.

The log entry carries the file name, who sent it and when it was read, followed by the summary. The email carries one line per document - the file name, the sender and what kind of document it is - and a link to the log, so the contents stay in the document you control rather than being forwarded around. Nothing is emailed when nothing arrived.

If the summary cannot be written, the log gets the document's own answers instead. Nothing is made up to fill the gap, and a document that could not be read at all gets no entry rather than an empty one - the run log says which.

A run reads at most 25 messages and summarises at most 'maxDocumentsPerRun' documents, which can be anything from 1 to 25.

Only PDF attachments are read; a signature logo is an image a document reader will happily describe, and it is not a document. Each message is labelled once its documents are logged, and the search skips anything already carrying that label, so nothing is summarised twice. The label goes on last: if a run dies part way the next one writes the entry again, which is a duplicate you can see, rather than skipping a document you never knew arrived.

New mail is what wakes it, but the sweep is what does the work, so several documents landing together are all read. Gmail only hands over the attachments of a sender's newest message in the inbox, so a second document from the same person in the same few minutes cannot be opened - the run log names it and says to forward it to yourself. For the same reason the mail must still be in the inbox when the run happens.

Reading a document is a smart action, so an occasional run pauses for review before it finishes.

Fill in 'documentSearch' with a Gmail search that finds the mail carrying documents, 'logDocumentUrl' with the Google Doc the summaries should be kept in, and 'notifyEmail' with your own address. 'processedLabel' and 'maxDocumentsPerRun' have working defaults.