AI-Assisted Genealogy Research: Ad-Hoc Prompting vs. Structured Frameworks

I've been dabbling in genealogy research for a few years now — mostly out of curiosity after my grandmother handed me a box of documents she'd been hoarding since the 1970s. When I started using AI tools to help parse old records and connect dots, I did what most developers do: I threw ad-hoc prompts at ChatGPT and Claude and hoped for the best.

It worked. Sort of. Until I realized I was asking the same questions five different ways, losing context between sessions, and had zero structure to show for weeks of research. Then I stumbled across autoresearch-genealogy, a structured framework built specifically for Claude Code, and it completely changed how I think about AI-assisted research workflows.

Let me walk you through why structured prompting matters, how these two approaches compare, and what it looks like to migrate from chaotic prompt-and-pray to something repeatable.

Why This Comparison Matters

Genealogy research is deceptively complex. You're not just asking an AI to summarize a Wikipedia page — you're cross-referencing census records, navigating digitized archives with inconsistent OCR, resolving name variations across decades, and building a knowledge graph of human relationships.

Ad-hoc prompting falls apart fast in this kind of work. Structured frameworks give you guardrails, templates, and — critically — a way to pick up where you left off.

The Ad-Hoc Approach: Quick but Fragile

Here's what most people start with. You open Claude, type something like this, and iterate:

markdown

# Typical ad-hoc genealogy prompt

I'm researching my great-grandfather John Kowalski who 
immigrated from Poland around 1905. He lived in Chicago.
Can you help me find records?

# Problems with this:
# - No structured output format
# - Context lost between sessions  
# - No systematic archive-by-archive approach
# - You'll forget which sources you already checked

You get a helpful response, sure. Claude might suggest checking Ellis Island records or the Cook County archives. But next week when you come back, you're starting from scratch. There's no research log, no template for tracking what you've found, and no systematic approach to working through archives.

I did this for months. I had a messy Google Doc with copy-pasted AI responses, half of which contradicted each other because I'd given slightly different context each time.

The Structured Approach: autoresearch-genealogy

The autoresearch-genealogy project takes a fundamentally different approach. It's built for Claude Code and provides three key components:

Structured prompts that guide AI through genealogy research methodically
Vault templates (think Obsidian-style) for organizing findings into linked knowledge
Archive guides that map out which records exist where and how to access them

Here's what a structured research session looks like in comparison:

markdown

# Structured research template example
---
subject: John Kowalski
born: ~1882
origin: Poland (region unknown)
arrival: ~1905
settled: Chicago, IL
status: researching
---

## Sources Checked
- [x] Ellis Island passenger records
- [ ] Castle Garden records (pre-1892, probably not relevant)
- [x] 1910 US Census — found entry, Ward 16
- [ ] Cook County naturalization records
- [ ] Polish parish records (need to identify village first)

## Findings
<!-- Linked notes with source citations -->

## Next Steps  
<!-- AI-generated based on gaps in research -->

The difference is night and day. You're not just chatting with an AI — you're building a persistent, structured research artifact.

Side-by-Side Comparison

| Feature | Ad-Hoc Prompting | Structured Framework |
|---|---|---|
| Setup time | None | 15-30 minutes |
| Context between sessions | Lost | Preserved in vault |
| Research reproducibility | Low | High |
| Archive coverage | Whatever you remember to ask | Systematic guides |
| Output format | Unstructured chat | Linked markdown notes |
| Collaboration | Copy-paste threads | Shared vault/repo |
| Learning curve | Minimal | Moderate |

The honest tradeoff: if you're doing a quick one-off lookup — "what year was Ellis Island opened?" — structured frameworks are overkill. They shine when you're doing sustained, multi-session research across dozens of sources.

Migrating Your Workflow

If you've been doing ad-hoc research and want to move to something structured, here's the practical migration path:

Step 1: Install Claude Code and clone the framework.

bash

# Get the structured research framework
git clone https://github.com/mattprusak/autoresearch-genealogy.git
cd autoresearch-genealogy

# The repo includes CLAUDE.md with project-specific 
# instructions that Claude Code picks up automatically

Step 2: Take your existing messy notes and feed them into the vault templates. This is the painful part — but you only do it once. Step 3: Use the archive guides to identify gaps in your research. This was the biggest win for me. I'd been ignoring entire record categories (naturalization papers, church records, city directories) simply because I didn't know they existed or how to access them.

Privacy Matters: Analytics for Your Research Site

If you're publishing your genealogy findings online — and many researchers do, to connect with distant relatives — you need to think about analytics carefully. Genealogy sites contain sensitive personal information, and you don't want Google Analytics harvesting data about who's searching for whom.

I've tested three privacy-focused alternatives:

Umami is my current pick for genealogy sites. It's fully self-hosted, which matters when you're dealing with family data. GDPR compliant out of the box, no cookies, and the dashboard is refreshingly simple. You deploy it once on a cheap VPS and forget about it.

javascript

// Umami setup — drop this in your layout
// Self-hosted means your visitors' data never leaves your server
<script
  async
  defer
  data-website-id="your-site-id"
  src="https://your-umami-instance.com/script.js"
></script>

Plausible is excellent if you don't want to self-host. It's a paid service with a clean UI and strong privacy defaults. Lighter than Umami in some ways, but you're trusting a third party with your traffic data. Fathom sits in a similar space to Plausible — privacy-first, paid, simple. Slightly more polished UI in my opinion, but also slightly pricier.

For a genealogy site specifically, I lean toward Umami because self-hosting gives you complete control. When your site tracks visits to pages about living or recently deceased people, that data sensitivity matters.

My Honest Take

Structured frameworks like autoresearch-genealogy aren't magic. They won't find records that don't exist online, and they won't replace the satisfaction of manually discovering a ship manifest with your ancestor's name on it.

What they do is prevent you from wasting time re-treading the same ground, help you systematically work through available archives, and give you a knowledge base that actually grows over time instead of being trapped in chat logs.

If you're just casually curious about a grandparent's birthplace, ad-hoc prompting is fine. If you're doing serious multi-generational research — especially across international records — a structured approach will save you dozens of hours.

The Claude Code integration is what makes this particular framework interesting. Having your research prompts, vault structure, and archive guides all living in a repo that Claude Code understands natively means the AI isn't just answering questions — it's working within a system designed for this exact use case.

Start messy if you want. Most of us do. But when you hit the point where you're Googling the same archive for the third time because you can't remember what you already found there, that's your signal to migrate to something structured.