A few months ago my dad handed me a stack of scanned PDFs — about a hundred so far, more coming, from a 2010 scan of his father's and grandfather's papers. Letters, naturalization documents, Greek refugee committee records, an Archdiocese recognition certificate, biographical newspaper profiles in Greek. Some typed, many handwritten. Names in Greek and in five different English transliterations of the same name across forty years of paperwork. Most pages were sideways because the scanner didn't care.
I'd inherited the thing every family eventually inherits: a box of documents nobody is ever going to read.
This is exactly the kind of work AI is for. Not "write me a poem." Not "build me an app." Just slow, careful, multilingual reading of a hundred documents, telling my great-grandfather Lazaros Jeannopoulos apart from the other Lazaroses across cousins, in-laws, and namesakes.
I built a small system inside a project I keep called homelife/ — a working notebook for the household, with a corner for ancestry. The flow is three steps.
1. Import. I drop a batch into /c/temp/lazaros/ and tell Claude /archive-import. It rotates the sideways pages, copies the scans into a durable archive, scaffolds an INDEX.md catalog (one row per file), and starts reading. PDFs in batches of four in parallel — slow enough to actually read them, fast enough to get through a hundred in an afternoon.
2. Remember. As Claude reads, it doesn't dump findings into the chat to be forgotten. It writes them into person-files: constantine_jeannopoulos.md, lazaros_jeannopoulos.md, eftyhia_jeannopoulos.md. One file per ancestor. Each accumulates documented facts, citations to the source document, the open questions still hanging, and a last_distilled date.
3. Distill. When a person-file changes materially, I run /distill-to-site and Claude generates a public-facing version for jeannopoulos.com/family. The public version drops the working-notes voice and keeps the durable facts. Anything I haven't double-sourced doesn't go up.
The findings are not what I expected.
I expected names and dates. I got names and dates. But documents have more in them than you know to ask. My great-grandfather Lazaros's naturalization date got pinned down by cross-referencing three independent papers. My grandfather Constantine — who I knew as a doctor in New York and not much else — turns out to have come through Cornell, made Phi Beta Kappa, then NYU medical school. My great-grandmother Eftyhia turns out to have been a Karamitrou from Pergamon, a fact that mattered last week because we're working on Greek citizenship and Pergamon puts her squarely in the 1922 Anatolian Greek diaspora and changes some of the paperwork we file.
I also found a thing I didn't want to find: a deletion entry in the Greek Mitroon Arrenon (the male register) for one of the chain, which is the kind of thing that can hand-grenade a citizenship case. It's now noted on the relevant page and we know what we're dealing with. Not finding it would have been worse.
There are two things AI did here that a human alone could not have done in any reasonable time.
The first is the language and transliteration mess. Greek documents from the 1920s through the 1950s have my great-grandmother as Ευτυχία, Eftyhia, Eftihia, Eftimia, Eftehia — different ratios across different archives. Claude tracks the variants, picks a primary, records the alternates, and matches them when the same person shows up in a different record under a different spelling. Doing that by hand across a hundred documents is the kind of work that ages you.
The second is the cross-referencing. A 1937 naturalization affidavit sets a date that contradicts a 1955 stamped form. Claude builds a hierarchy — priest-signed primaries over civil re-issuances over US-records derivatives — and the page records both the winning value and the reconciliation behind it. Future me, or my daughter Mia in twenty years, will be able to see why a date is what it is.
What's at jeannopoulos.com/family today is a working record. Five generations, two main branches — the Greek-Anatolian Jeannopouloses and, through my wife Karyn, the Lebrun-Chassaing line out of Port-au-Prince. About twenty-five people so far, more to come. Each person has their own page. Each page has open questions still on it, because that's an honest archive.
When people ask me what AI is good for, I send them this. It's not a magic trick. It's not "the AI did my genealogy for me." I read every page Claude wrote. I argued with it about the date of one document for half an hour because two sources disagreed and the disagreement turned out to mean something. The AI is doing the patient, multilingual, repetitive reading I would never have gotten to in this lifetime, and putting it into a shape I can think with.
That's the gist of it.
If you've got a box of family papers and you're staring at it the same way I was, I teach this kind of practical Claude work in small in-person classes through Gist in Broward and Palm Beach. Book a free intro call or email [email protected].
— Alex