How I Stopped Being Afraid of Family Archives and Delegated Genealogy to Claude Code

In many of our closets sits that very box. Inside it — yellowed photos of unknown relatives, birth certificates in an unfamiliar language, and possibly a couple of letters from the front. We promise ourselves we'll "eventually sort through all this," but the volume of data is intimidating. Genealogy isn't just drawing a tree — it's an endless cycle of archive searches, hypothesis testing, and wrestling with contradictions in documents.

A project called autoresearch-genealogy has appeared on GitHub, bringing autonomous agent methods to the realm of family history. Its creator, Matt Prusac, didn't just dump a set of prompts. He systematized real research experience during which AI helped reconstruct 9 generations across 6 family lines, creating over a hundred detailed files in the process.

What's inside the repository

The project is a ready-made methodology for working with Claude Code or any tool that supports autonomous searching. The main value here isn't in the code — it's in the process structure.

Prompts for autonomous searching

Folder prompts/ contains 12 scenarios for the /autoresearch command. Each prompt is a mini-program for AI. It specifies objectives, success metrics, verification protocols, and "limiters" to prevent the neural network from hallucinating.

For example, prompt 01-tree-expansion forces the model to methodically "grow" each branch of the tree using web search. And 05-source-citation-audit takes on the role of a strict archivist: it verifies that every person in your database has at least two independent sources.

Digital repository template

The author proposes using Obsidian (or any markdown editor) as a database. Folder vault-template/ contains the structure for your future archive:

• Person cards with YAML metadata.
• Research logs and registries of open questions.
• Special templates for transcribing documents, letters, and even DNA match analysis.

Guides on archives and methodology

This is perhaps the most practical part. The repository includes instructions for searching in 24 regions — from the US and colonial records to archives in Poland, Ukraine, and Russia. There are even sections on Jewish genealogy and the specifics of Scandinavian surnames.

Why this works better than a regular AI chat

The main problem when using ChatGPT or Claude for such tasks is the models' tendency to "fill in" facts. Genealogy can't tolerate this. If one document records a person as "Zachariah" and another as "Sakkarias," the AI shouldn't just pick one variant — it should record both and assess the probability of error.

Matt Prusac embedded a "mechanical verification" concept into the prompts. This is similar to a compiler's work in programming, but instead of checking syntax, it verifies the logic of relationships. The system forces the AI to:

1. Log negative results (what was searched for and not found).
2. Assign confidence levels to facts (Strong Signal / Speculative).
3. Perform cross-reference audits after each search iteration.

How to try it with your own data

If you have Claude Code installed, the process is fairly straightforward. First, you need to clone the repository and move the contents of vault-template/ into your working directory.

Then you enter basic data into the Family_Tree.md file — what you already know about yourself, parents, and grandparents. If you have document scans, place them nearby. After that, you can launch the first cycle: copy the contents of prompts/01-tree-expansion.md into the terminal with Claude and watch the agent start "digging" the internet for leads.

Who this will help

The project is unlikely to suit those who just want to press a button and "learn everything about their ancestors." Genealogy is still hard work. But autoresearch-genealogy will be an excellent resource for:

• Techies who find filling in tables manually boring but are interested in setting up an autonomous pipeline.
• Researchers stuck on a "brick wall" who need a fresh perspective on available archives.
• Owners of large volumes of scattered data that need to be brought to a unified standard.

Interestingly, the author doesn't restrict usage to Claude Code only. The entire logic easily transfers to a manual workflow or other AI tools. The key here is the confidence level system and verification protocols that transform genealogy from reading tea leaves into structured research.

If you've long wanted to turn that very box of photos into a coherent family story, this repository might be the missing push you needed. At minimum, it's a great example of how to use modern LLMs for solving complex humanitarian problems without losing scholarly rigor.