Claude Code traced the author's ancestors back to 1728 in a week — a genealogist would have charged from 230,000 rubles
The author set out to check his grandfather's story about ancestors who were industrialists, and he asked Claude Code to read scans of archival parish registers. In one week, the agent assembled documents covering seven generations. The earliest ancestor was born around 1728 and was a serf at the ironworks. Genealogists charge from 230,000 rubles for a branch like this; the author only needed the subscription he already paid for. The takeaway: routine archival search can be handed to an agent, provided the fact-checking rules are spelled out in advance.
AI-processed from Habr AI; edited by Hamidun News
The author's grandfather once told him, during a taxi ride, that the family were hereditary industrialists, practically partners of the Demidovs, until the Bolsheviks came and took everything. The grandson did not believe it, but the idea stuck, and in the autumn he set out to test the legend together with an AI. His partner was Claude Code, an agent that lives in an ordinary folder on a computer, can see the files, and remembers where the work left off.
Over the course of a week, the chain of archival documents covered seven generations, and the earliest ancestor found, Timofey Shibakin, was born around 1728. According to the author, professional genealogists charge from 230,000 rubles for a single branch like this. He needed only a week and a subscription that was already paid for.
How the work with the agent was organized
The author did not start with the AI but with a survey of relatives: without a lead such as a village or district, the agent is useless. Half the answers went something like "somewhere near Kungur, I think," but that was enough to determine which archives to visit. Three of the four surnames belonged to Perm Governorate before the revolution, so most of the work went through the archives of Perm and Yekaterinburg.
The project folder holds the rules (CLAUDE.md), a database of people with a unique number for each, a search plan, transcriptions from the archives with references, letters, and scans. The most important part is the set of rules that applies in every session:
- each fact has a source and a status: document, "according to relatives," or hypothesis;
- a hypothesis may not be presented as a fact, and a namesake from the database is not counted as a relative until the place and dates match;
- letters to archives go out only after the author's explicit "yes";
- personal data of living people is not published externally.
The work was divided as follows. The agent read scans of parish registers and revision lists in online viewers: 19th-century handwriting, pre-reform orthography, abbreviations, torn pages. It checked dates, maintained the database, looked up case numbers, and drafted letters. The human interviewed relatives, logged into portals via Gosuslugi (the agent does not enter passwords), sent the letters, and made the decisions.
The author also found a ready-made skill, ex-ussr-kinship-research, on GitHub, but did not install it in full: there, the agent solves the CAPTCHA itself and sends already approved letters without asking for confirmation again. He did take its techniques, though. For example, a negative result is recorded together with its coverage: which parish, which years, which sheets. "Not found in the checked records" and "did not exist" are different things.
Where the rules saved the branch
A telling episode was the search for the great-grandfather, Nikolai Shibakin. According to the documents, he was born on 27 October 1904 in Pervouralsk. In the parish register, a Nikolai did appear on that date, but he was the son of a reserve private named Portnov. The agent reread the whole year and found entry No. 135: Nikolai was born on 12 August, and his father was the peasant Aleksandr Gavrilov Shibakin. In the later documents, the date had simply drifted by more than two months.
There was also a trap: the same book contains another Nikolai Shibakin, born in April, whose father is also named Aleksandr, though with a different patronymic. Had the agent taken the first Nikolai it came across, the entire branch would have run off to strangers. This is precisely why the author wrote the rule about namesakes.
From there the chain went upward to a 1811 revision list that records "Prokhor, son of Sofron, colloquially Shibakin." At that time there was no surname, only a street nickname. In the neighboring yard was Timofey Gavrilov, 67 years old according to the 1795 revision, who died in 1796, which puts his birth around 1728. Whether he was Prokhor's grandfather or great-uncle remains a hypothesis. Documentation establishes a direct line only as far as Sofron, Prokhor's father.
The outcome for the grandfather's legend: the ancestors were serf workers at the ironworks, not industrialists. But the Shaitan Works, where they lived, were founded by Demidov, so "under the Demidovs" was correct, just on the other side of the dam. The name of the later owner of the works, Berg, the grandfather also got right. On the paternal line, a grim family history was confirmed as well: in the database of victims of political repression, the author found cards for fellow villagers named as participants in the 1918 Yurla uprising and shot in 1937.
What this means for practice
The case is interesting not for the genealogy but for the way the process was built. The agent was given file-based memory instead of hope that context would carry over, strict status labels for reliability, and a ban on unilateral external actions. The human kept the decisions, sending the letters, and logging into the portals.
There are honest limitations too. Many questions remain open: the inquiry about repressed relatives is sitting with the Perm archive, and the link between Timofey and Prokhor, along with part of the versions about the author's father, are still marked as hypotheses. The agent speeds up the routine of reading manuscripts, but it does not replace verifying the conclusions.
For those building agents for long research tasks, the lesson is simple: first spell out the rules and fact statuses, and only then let the model loose in the archives, documents, or knowledge bases.
- Nov 21, 2026FDE job postings grew 4.6x in a year: what a forward deployed engineer actually does at the client's site
- Nov 21, 2026Anthropic cuts Claude Fable 5.2 price by 20%, and on the coding test it nearly matches Opus 5.5
- Nov 21, 2026Dreamteam Raises $53M for a CRM That Fills Itself from Email and Calls
- Nov 21, 2026Claude Opus 5.5 Builds Out Dimension 9 Horror Game From the Author's 15 Rooms
Key Questions Answered
How the work with the agent was organized?
The author did not start with the AI but with a survey of relatives: without a lead such as a village or district, the agent is useless. Half the answers went something like "somewhere near Kungur, I think," but that was enough to determine which archives to visit. Three of the four surnames belonged to Perm Governorate before the revolution, so most of the work went through the archives of Perm and Yekaterinburg.
Where the rules saved the branch?
A telling episode was the search for the great-grandfather, Nikolai Shibakin. According to the documents, he was born on 27 October 1904 in Pervouralsk. In the parish register, a Nikolai did appear on that date, but he was the son of a reserve private named Portnov. The agent reread the whole year and found entry No. 135: Nikolai was born on 12 August, and his father was the peasant Aleksandr Gavrilov Shibakin. In the later documents, the date had simply drifted by more than two months.
What this means for practice?
The case is interesting not for the genealogy but for the way the process was built. The agent was given file-based memory instead of hope that context would carry over, strict status labels for reliability, and a ban on unilateral external actions. The human kept the decisions, sending the letters, and logging into the portals. There are honest limitations too.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.