Everything you can hand the book agent: photos, messages, videos and a shoebox
Part two of five in the series A book with the book agent. Previous: A book about your children’s childhood: start with your voice, not the keyboard. Next: A voice note from the car.
In part one we said the whole method rests on two things: the voice note and the material you hand the agent. This part is about the second. The short answer to the question in the title: almost anything with words or dates in it. The long answer is below.
A model example
Jana (our model client from part one – not a real person, an example assembled from common situations) spends her first week listing what she actually has:
- about 11,000 photos and 40 videos on her phone and in the cloud, mostly without captions,
- a family group chat with the grandparents that has been running since 2017,
- a baby book she wrote in until month fourteen and then stopped,
- e-mails she sent to her mother abroad when the children were small,
- nursery portfolios – on paper – and a shoebox of drawings with a date pencilled on the back,
- the calendar on her phone: first tooth, first day of school, vaccinations, visits.
Until that moment Jana thought she “had nothing to write from”. She has enough for three books. It had just never been in one place.
What the agent reads
Documents. PDF, Word, plain text, saved web pages, spreadsheets, presentations. Whole books in EPUB too – those are useful for the style analysis described in part one.
Message archives. Exports from WhatsApp and Telegram, a mailbox, an SMS backup. This input gets its own section below, because it is the most valuable and the most underrated.
Images. Photos, screenshots, scans. The agent does not just store them – it reads them. Give it a photographed drawing with a date on the back and it reads the date. Give it a photographed note from the teacher and it reads what it says. Give it a photo of a room and it tells you what is in it.
Audio. Voice notes of any length – five minutes, thirty, more. Recorded conversations, an old dictaphone, the children’s voice memos. Long recordings are transcribed in parts.
Video. The audio is extracted and transcribed. For what three-year-old Tomáš says while blowing out the candles, that is all you need.
Web pages. Send a link and the agent reads the page.
How it files it: four layers
This is the reason the book holds together in the end, so it is worth a paragraph.
1 · Raw material. Originals are never altered. Searchable text is extracted from each document, with a note beside it: which dates were found in it, which names, how big it is. A year from now, when you are no longer sure what exactly that letter from the nursery said, the untouched original is there.
2 · Your voice. Each voice note becomes one file: the verbatim transcript at the top, the analysis below it, the open questions at the end. Nothing is paraphrased. In a true story the exact words are evidence – and they are also what your voice in the book is later built from.
3 · Facts. Every fact is one line: a number, a date, what happened, how we know, and a status tag. Part four is all about the tags; here it is enough to know that a fact which cannot be traced to a document or to your own words does not get into the book.
4 · Decisions. What the book is about, what shape it has, how it sounds, what is settled and will not be reopened, what you still owe. Decisions are written down so they are made once. Without that, every session argues the same three questions again and the book never starts.
Send it twice and nothing happens
A parent with eleven thousand photos cannot remember what has already been sent. There is no need to. When the agent receives a file it already has, it says so – under which number it is filed – and sets the copy aside. Nothing is duplicated, nothing is lost.
The message archive: the most underrated input
A family group chat running since 2017 is probably the most valuable thing Jana has – and the last thing she would think of handing over.
It holds dates nobody remembers. The exact words grandma wrote on the day Tomáš came home from the maternity ward. A photo with the time it was sent. A sentence from December 2019 that means more today than it did then. Once handed over, years of messages can be searched by date, by word – and by silence: when nothing was written for three weeks, and why.
In the real project on which we developed the method, it was an export of 13,181 messages over 45 months. In it the agent found an agreement the author had forgotten and corrected a date the author had been sure of for two and a half years. Those numbers come from that project, not from Jana’s story – but the principle is the same for a group chat with the grandparents.
One honest limit: scanned PDFs
A PDF that is really a photograph of a page contains no text – only a picture of text. You can tell in five seconds: open it and try to select a word with the mouse. If it highlights, it is fine. If nothing selects, it is a scan.
The fix is simple: open it in Word, which extracts the text, and save it as .docx. Or photograph the page and send it as an image – the agent reads images. We say this up front, because a page that admits no limit cannot be taken seriously about anything.
Where it all lives
On your computer. The children’s photos, the messages with the grandparents and the notes from the nursery are not uploaded anywhere and are not used to train any model. The one precise exception that has to be said out loud: the language model itself is an external service, and the agent sends it what it is processing at that moment. The material stays at home.
What comes next
Part three is about what exactly happens to one thirty-minute voice note between the moment you send it from the car and the moment you park: A voice note from the car: what happens to 30 minutes of talking.
Want to see it live
See how the book agent works – it is all on one page. If you want to watch the agent read a photographed note from the nursery, book a free 30-minute call with a demo. A live demo also runs in our Discord community.