How to Use Claude Code for Deep Research Projects Like Genealogy

A developer recently demonstrated a powerful application of Claude Code beyond typical software tasks: automating deep, evidence-based research. By combining Claude Code with a specialized agent framework, they turned a sparse family tree into a detailed, source-backed genealogy project tracing back to Jamestown.

The Technique — Agentic Research with Claude Code

The core method was using autoresearch-genealogy, a project inspired by Andrej Karpathy's "autoresearch" concept. This isn't a single tool but a framework of prompts, guidelines, and guardrails that instructs Claude Code to act as a research agent. The agent is given a clear metric to optimize and a set of tools (like web search and data parsing), then runs for a minimum number of iterations to gather and synthesize information on a scoped topic.

The project provides a templated Obsidian vault structure. As Claude Code runs, it populates this vault with markdown files—one per person—linking individuals, embedding sources (census records, death certificates, newspaper clippings), and explicitly noting conflicts or unknowns. The result is a graph of evidence, not just a flat family tree.

Why It Works — Claude Code's Strengths for Structured Tasks

This approach leverages Claude Code's fundamental strengths: following complex instructions, using external tools, and producing structured output. The agent framework provides the necessary guardrails to keep the research on track, while Claude Code's ability to reason about relationships and synthesize disparate data points turns a manual detective process into an automated one.

Crucially, the agent is instructed to preserve ambiguity. It flags single-source claims, logs failed searches, and maintains conflicting accounts instead of flattening them. This creates a transparent audit trail and highlights where human judgment is still needed, making the automation assistive rather than blindly authoritative.

How To Apply It — Building Your Own Research Agent

You can adapt this pattern for any deep-dive research project (market analysis, academic literature review, competitive landscaping). Here’s the workflow:

1. Define Your Scope and Metric: Start with a clear, bounded question and a success metric (e.g., "Find 20 primary sources on Topic X" or "Map the ecosystem of competing tools").
2. Set Up Your Knowledge Vault: Create a simple directory structure. The genealogy project used an Obsidian vault, but you can start with a plain research/ folder containing a CLAUDE.md file.

   <!-- CLAUDE.md for a Research Project -->
   # Project: Competitive Analysis for LLM Code Tools
   ## Goal
   Identify and categorize the top 10 competitors to Claude Code, detailing their key features, pricing, and differentiators.
   ## Output Structure
   - Create one markdown file per competitor in `/competitors/`.
   - Each file must include: description, core features, pricing model, and a "vs. Claude Code" section.
   - Maintain a master `INDEX.md` with a comparison table.
   ## Research Rules
   - Prioritize official documentation and recent (2025+) articles.
   - Note any conflicting information about features.
   - Log sources for every claim.
   

3. Let Claude Code Run Iteratively: Use the claude code command in your project directory. The agent will use its tools (like web search via MCP) to gather data, synthesize it, and populate your vault according to the CLAUDE.md rules. You can run it in multiple sessions, each time refining the prompt based on what's been found.
4. Build on the Data: Once you have a structured knowledge base, you can direct Claude Code to build on it. In the genealogy example, the next step was: "Now, build a simple Next.js app that lets me browse these people and visualize the relationships." Claude Code then generated the application code using the markdown files as its data source.

The Takeaway: From Research to Application

The final step showcases the full power of this workflow. Dissatisfied with existing genealogy visualization libraries, the developer simply told Claude Code to "make your own." It proceeded to write an SVG generator from scratch to visualize the entire family tree. The structured research vault provided the clean data necessary for this custom tooling.

This case study proves Claude Code can be the engine for end-to-end knowledge projects: from open-ended research and evidence collection to structuring that knowledge and finally building custom software interfaces to interact with it. The key is providing the right guardrails and a clear structure for the output.