Image: GitHub
Why it mattersThe skill labels each claim with a plain status like 'corroborated' or 'single-source'. Any team building AI products can copy that design.
Finding a great-grandparent in an 1890s Ukrainian church book is slow work. The researcher first needs to know which parish served the village. A new AI agent skill tries to do that work itself. It has drawn 2,394 GitHub stars in the 19 days since it was pushed.
The skill is called ex-USSR Kinship Research. It was published on 25 October 2026 by Denis Shiryaev under the MIT licence. It covers all 15 former Soviet republics and the Russian Empire territories that preceded them. The list runs from Estonia to Uzbekistan. Shiryaev writes that it reads church books, civil records and military files. It also handles deportation and repression files, and scans of unindexed archive pages. It reads FamilySearch indexed records and existing GEDCOM family trees. It covers First World War hospital, award and prisoner records as well.
Six evidence statuses
Most AI tools give an answer with a confidence score. This skill uses plain text labels instead. Each claim gets one of six statuses: hypothesis, single-source, corroborated, strongly-corroborated, contradicted, or unresolved. A copy of a record and the original share an independence group. Two copies of the same scan count as a single source. An audit script runs before any conclusions are written. It checks that every referenced artifact and person exists. Shiryaev writes that "corroboration cannot be declared into existence".
The skill's instructions are written in Russian. The agent replies in the user's own language. The README lists English, Ukrainian, Russian, Polish and German as tested languages. The README itself is published in more than 20 languages.
What the skill refuses to do
The README lists several hard limits. It refuses to look up current addresses or accounts of living people. Shiryaev describes the skill as a tool for family history research only. It refuses to upload documents or DNA data to third parties without consent. It makes one CAPTCHA attempt and then stops. It refuses to bypass paywalls or evade rate limits. Sensitive findings about deportation, repression or adoption are reported neutrally. They are reported only as far as the research question requires.
From a case folder to a PDF
A user creates a case folder, naming the research question and the subject. The user drops scans of documents into the folder. The skill fingerprints each one. The agent then reads its registers of persons, artifacts, claims and relationships. It does the research against public archives. An audit script runs the integrity checks before conclusions are written. A build step turns the final report into a PDF for family members who do not read agent output.
The README names Codex and Claude Code as supported runtimes, and gives one-line install paths for each. It also supports the general Agent Skills CLI.
The slow part of family history research is knowing which record set to open first. That is the part the skill tries to take over.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.