Codewalkers

One Evening, Start to Finish – A Real Genealogy Session

·
9 min read

I’ve written about how I set up Claude Code for genealogy, the admin section behind the site and how it changed my workflow. All of that is the how. This post is one real session, from Sunday evening into Monday morning, with the actual steps, findings and mistakes.

The whole thing took about a dozen messages from me. Most of them were short. “Run jobs 16 18 and 19.” “Go ahead.” “Update?”

Getting started

Every session starts with /start-session, and it runs the same steps every time. First it pulls in whatever I’ve sent from the admin site, like answers to to-dos and headstone photos. Then it asks whether I want to refresh the tree from FamilySearch. I said yes. A full download takes about 30 minutes, so it runs in the background, and the new copy only gets swapped in if it passes its checks. This one came back with 1,533 people, 726 families and 16,601 sources.

Then it asks me three things. Do I want to add anyone to scope? (No.) Do I have a burning question? (Yes. Who were the parents of James M. Wade, my third great-grandfather?) And which research jobs should it run? Those come from a backlog of work it can do without me, each with its size and cost. I started with three, and by the end it had run seven and was about halfway through an eighth.

The work gets split up by family. Eighteen research agents ran in parallel groups, each assigned one branch of the family. Each one writes only its own files, so no two agents ever edit the same thing. Anything outside its lane comes back as a proposal for the main session to review. Eighteen agents, no collisions.

One finding, start to finish

The easiest way to show how it works is to follow one finding all the way through.

Charles S. Walden is a great-great-granduncle on the Webb side. Before this session, the tree had no dates or places for him at all.

FamilySearch has Full-Text Search, which machine-reads old handwritten court books and makes them searchable by name. It had text suggesting a “C. S. Walden, Sheriff” of Burleson County, Texas. That’s a lead, not a fact. Machine-read text never goes on the site as a fact.

So Claude Code opened the page images and read them itself. There were two writs from 1850 endorsed “C. S. Walden Sheriff B.C.” There was a writ from 1848 that “Came to hand Augt 22d 1848,” signed by Charles S. Walden as sheriff of Burleson County. And there was a sheriff’s deed dated the 3rd of February 1852.

Each page became a document record: a citation, a transcription, a table comparing what the document says with what the tree says, and a classification of the evidence. Then comes the step I always skipped when I did this by hand, the follow-through. A script lists everything the new record could touch (open questions, to-dos, earlier facts that disagree, inferences in the biography) and each one gets an outcome. The project won’t let a new record be committed without it. In this case, a sentence in his biography had been marked as inference because it rested on the machine-read text. It got replaced with a cited fact.

Then it wrote me a to-do. It tells me exactly what to add in FamilySearch: the occupation “Sheriff of Burleson County,” dated 1848–1852, with the reason text and the image links ready to paste. It also posted a discovery to the site, and amended it twice during the session as the dates widened from 1850, to 1848–1850, to 1848–1852.

So a man with no dates in the tree is now a county sheriff for at least three and a half years. Every word of that cites a page that was actually read.

What else turned up

The sheriff wasn’t the only one.

  • Harris McKinney’s will (1850) and codicil (1855). The will names five children, including a daughter, Mary Chappel, who wasn’t in the tree. It also shows that two people the tree gave him as children weren’t his. The codicil names a granddaughter, Celia McAdory, who wasn’t in the tree either. County histories say she was born in November 1837, a month before her father died. The codicil is the first original record found that names her.
  • Charles Webb’s grave (1849). When his executors sold the home tract, they kept back “twenty feet square around the grave where Charles Webb is buried.” A family grave on private land, located by a deed. That closed a search item that had been open a long time.
  • Charles Walden’s estate (1834–1835). The Orphans Court named heirs that included two sons the tree didn’t have, and gave the father’s middle name, Sparks.
  • Edmund Carlisle’s will (1828). Nine children, and a family plan to move west “to the Chatahoochy.” It also explained an error. The death date in the tree was really the index’s reading of the date he wrote the will.
  • John Wade, silk dyer, 74 Gold Street, New York. City directories from 1800 to 1820 and an 1807 naturalization tied the Gold Street Wades to the family behind my burning question.
  • Alban Willingham’s heirs (1858). A widow’s renunciation of dower named his married daughters and their husbands.

Some of these records are hard reading. Wills and estate sales from that era list enslaved people as property. McKinney’s codicil gave his son Patrick “my faithful servant Girl named Eliza & her three children.” At Sampson Ivy’s estate sale in 1816, ten people were sold apart from one another. The project writes these up plainly, the way the documents give them, and when a record names the people, so does the write-up. Often that record is the only place their names survive.

The burning question

This one didn’t get answered. I didn’t expect it to.

The session took it as far as it could on its own. It went through the Manhattan death registers for 1824 to 1833 (27 Wade deaths), checked the 1820 and 1830 censuses for the city’s Wades, ran two Full-Text searches, and read three histories of the Baptist church the family attended.

The tree’s candidate father, Edward Wade, doesn’t head a household in New York City in 1820. That adds weight to a doubt that was already on file. It wrote everything up in a dossier with the leads and a costed plan for stage two: a group of agents reading about 40 pages of text and 20 to 25 page images. If that doesn’t crack it, stage three is a brief for a professional genealogist. Neither one runs until I say so.

A hard question doesn’t get answered in one sitting. It gets narrowed, written down, and priced out. That’s still more than I used to get out of an evening.

What went wrong

It made mistakes. Honestly, this is the part I find most useful, because each one either turned into a rule or got caught by a check.

It got turned away by FamilySearch. The project is deliberately polite to its sources. Page images are read one at a time, at most 20 new ones a day, never in a loop, and everything stops at the first refusal. Partway through, it asked for “the next page” instead of a page’s direct address. That triggers a lookup FamilySearch refused with a 503. It stopped, told me, and explained why. I told it to resume using direct addresses only. That lookup was already a known problem with a fix proposed, so this just confirmed it.

The index was wrong. Three times. FamilySearch indexes Sampson Ivy’s guardian bond as 1846. On the page it’s clearly 1816. Someone read the loop of a 1 as a 4. A to-do that relied on the index got corrected before I acted on it. The same session found an indexed “Thomas” that was “Shimie” on the page, plus the Carlisle will date that had been sitting in the tree as a death date.

The shell ate a dollar sign. A log row lost “$745” because the shell treated $745 as a variable. Its own review caught it, and the quoting rules got tightened.

It contradicted itself. The task prompts it wrote for its agents named the wrong logging command, which disagreed with the agents’ standing instructions. That became a tool proposal, so the standing instructions stay the only source of the agents’ rules.

It kept stopping to check in. During the image reading it kept pausing to give me status reports when nothing needed my decision. So I asked, “Are you waiting on me on job 026? If so, why? Why don’t you just keep doing it and finish it?” From then on it read straight through to the daily limit. Shortest instruction I gave all session, and probably the most useful.

Where it left things

By Monday morning:

  • 19 new document records (wills, deeds, court writs, an estate sale, a guardian’s bond, a census page, city directories)
  • 72 biographies revised and 9 new discoveries on the site
  • 544 rows added to research logs across 133 people, including the searches that found nothing
  • 13 new to-dos for me, plus dozens of existing ones corrected or sharpened
  • 50 proposals for improving the tools
  • 12 commits, all pushed

A lot of that is now my job. The to-dos tell me exactly what to change in FamilySearch: add Mary Chappel, add Celia McAdory, add the sheriff’s occupation. I make every one of those changes myself, and Claude Code never writes to FamilySearch. That’s on purpose. It’s how I keep the story in my own head instead of just in a database.

One courthouse trip got smaller, too. I had a to-do to visit the Coosa County probate office. The will and codicil turned out to be online, so now I only need to ask for the inventory at the counter.

Claude Code has the rest of job 026 waiting for tomorrow’s image allowance, about 20 more pages of heirs’ papers, estate settlements and deeds. About 370 General Land Office patents have been judged but can’t become records until there’s a tool for that source. The 50 tool proposals wait for me to accept or dismiss them in a separate technical phase. Research sessions do research. Building tools is a different job, and keeping them apart keeps the sessions from wandering off.

What I took away

The thing I keep coming back to is how little of this I did. About a dozen messages. But nothing in FamilySearch changed without me, and every claim on the site rests on a page that was read. The indexes were wrong more often than I’d have guessed. And the process remembered everything, dead ends included, so the next session picks up exactly where this one stopped.

That was always the part I was bad at.

Everything from this session is on webbwade.com.

Say something

Your email won’t be published. Name and email are required; the comment is the important part.