In a useful oversimplification, there are two types of people:
Those who learn by reading the frickin’ manuals
Those who learn by pushing buttons
In the coming days, weeks, and months, there will be great weeping and gnashing of teeth among the experiential learners--the “button-pushers.” There are new AI tools available that are going to empower both of these types of people to do work in new and powerful ways. And like chainsaws, dynamite, and bulldozers, tools can be both used and misused.
Button-pushers need to be aware that these tools can do real damage when used carelessly or recklessly, or even merely inattentively. Misusing these tools can result in the loss of data and the loss of work, even the lost of your hard drive or cloud storage. Experienced programmers have turned their computers into bricks, paperweights, and deleted their life’s work by falling asleep at the wheel, or by not attending closely to what these tools are doing. So, everyone, both button-pushers and manual readers, need to be aware of the risks and take steps to mitigate them.
Vibe Coding becomes Vibe Work becomes Vibe Genealogy
In February 2025, AI researcher Andrej Karpathy coined the phrase “vibe coding.” This was two years after he foresaw that English would be the hottest new programming language. And for some time now, I have been echoing, “Soon, people will be able to speak tools into being.” Well, that time is now.
Vibe coding is the term coined by Andrej Karpathy to describe a way that computer programmers—or folks who would like to be computer programmers—can use new AI-enhanced tools to accomplish real work rapidly without great expertise or training. There are now ways to use these tools that are so much easier and accessible that they are no longer limited to computer programmers. Indeed, vibe coding is expanding into “vibe work” where people without any computer programming training or background can use these tools for basic knowledge work, traditional white-collar work. Including genealogy.
Karpathy coined the term in a positive sense, but detractors and deriders sometimes use the term in a dismissive or insulting manner. This is to be expected. I wear another hat in another part of my work, and I am familiar with how language is used this way. The term “Methodist” was originally coined as a slur, a dismissive term, an insult, meant to belittle others, the followers of John and Charles Wesley who had discovered within the Anglican tradition a path, a methodology, to encounter the sacred. Rather than fighting against the term, however, the Wesley brothers and their followers leaned into it and embraced the name Methodist. The same will happen with vibe coding, vibe work, and vibe genealogy. There will be gatekeepers and pearl-clutchers who will try to use the term in a dismissive or insulting manner, but the term will be embraced by those who use these tools for real work. And real work will be done, done well, done with joy, and “vibe” will prevail in the fullness of time over the objection of folks who will not take the time and do the work to evaluate the tools and use them for real work.
I spent the month of December exploring and documenting the use of these tools and have embraced the phrase “vibe genealogy.” The advent of two things in late 2025 precipitated a second phase shift in the AI revolution, which began in November of 2022. The first phase shift began with the release of GPT-4 in March of 2023. If you were around and active then, you know that that was a significant moment. And more recently, in late 2025, the release of Anthropic’s Claude Opus 4.5 and their agentic tools have precipitated another phase change. It may not yet be visible to everyone except those who use these tools extensively, but it is now clear to those who do that we’ve entered a new phase in the AI revolution with the release of Claude Opus 4.5 and Claude Code. To those with eyes to see and ears to hear, something significant has changed since fall of 2025.
Prof Ethan Mollick, a Wharton professor and a respected voice in the field, has recently written about this (“Claude Code and What Comes Next,” 7 Jan 2026). He built a complete startup website in 74 minutes using Claude Code, with no coding knowledge required. That project for Mollick was a way of illustrating for all of us that agentic tools beyond coding are imminent. The harness that makes Claude Code work autonomously for programming tasks has extended to other knowledge work: data analysis, research, design, and yes, genealogy. Mollick takes care to explain that this capability shift is happening now, not later. Professionals who wait to experiment will develop what Mollick calls a “skill issue,” falling behind competitors who integrate these tools now. He also takes great care to emphasize the warnings with which I began this post. Mollick warns about accidental file deletion, unintended code execution, and unauthorized data access. His practical advice is this: “make backups, use a dedicated folder, and do not give it access to anything you cannot afford to lose.” In this post, I will demonstrate all of these things, especially the use of dedicated folders as projects.
Vibe Genealogy Assistant
In observance of my 59th circuit around the sun, I spent yesterday doing what I love - creating, testing, writing about, and sharing genealogy tools with fellow family historians. And in the tradition of Fun Prompt Fridays, which I have been doing for some time, I’d like to share with you a way to get started in vibe genealogy with a saved prompt. Depending on your skill level, the following saved prompt could be used one of several ways depending on whether you’re a beginner, an intermediate, or a power-user of these tools.
Beginners could use the prompt below in a one-off chat with ChatGPT, Claude, or Gemini. You might attach a small GEDCOM file and the following prompt, and ask the model to process the GEDCOM file according to the prompt, as a stand-alone set of instructions to process the GEDCOM file in chatbot. For this use, you can just copy the prompt into Claude (preferably Opus 4.5), ChatGPT (preferably GPT-5.2), or Gemini (preferably Gemini 3 Pro), and upload or attach a small GEDCOM file (remember, for example, that Claude has a context window of about 200 kb). Start a conversation with the chatbot about the GEDCOM file. In addition to GEDCOM files, you could also process any sort of meaningful genealogical source, such as record images and documents.
Intermediate users could use the prompt below to power a Project at ChatGPT or Claude, to power a Gemini Gem, or as a resource in NotebookLM. Intermediate users will know about Projects and how to create projects and Custom GPTs or as they’re called at Google’s Gemini, Gems. In these AI workspaces, the prompt is saved as custom instructions so that the user can then rely upon to inform and power that AI assistant. I’ve created several of these that you can try for free. For example, you can use this prompt as a Custom GPT at OpenAI or as a Gem at Gemini. In NotebookLM, you can save the prompt as a Source item and have NotebookLM reference and run that prompt from the Sources. Each of these workspaces could then be populated with source documents, images, and records to be processed by the prompt.
But the primary reason I share this prompt here today is so that advanced and power-users can try this prompt to make their first experimenting with Claude Code and vibe genealogy more focused, productive, and meaningful. I will discuss the ways to access and use Claude Code in its various incarnations in more detail below, but here, first, is the prompt:
BEGIN PROMPT (copy everything inside the code block below)
# Vibe Genealogy Assistant v4
You're a warm, knowledgeable genealogist who helps everyday family historians understand their heritage. You combine rigorous methodology with genuine curiosity about the families you encounter. Behind every date is a life. Behind every gap is a story. Your job is to help people see their ancestors as *people*, not just names.
## Your Three Modes
Adapt based on what the user provides:
1. **GEDCOM file uploaded** → Lead with story-first analysis, find the human meaning
2. **Document image uploaded** → Extract data, explain context, suggest next steps
3. **Research question asked** → Provide GPS-informed guidance at the user's level
Always adapt to the user's apparent experience—warm encouragement for beginners, collegial collaboration for intermediates, peer-level precision for advanced researchers. Never ask "what's your level?"
## Sacred Trust
Family history is sacred trust. You handle it with care:
**Privacy**: Never share identifying details of living persons without explicit permission. Living = born within last 100 years without confirmed death. When in doubt, protect.
**Sensitivity**: Trees reveal secrets—unknown parentage, hidden adoptions, children who died young, marriages family stories erased. Don't dramatize. Don't hide. Name what the data shows gently, and let the user lead: "I see some complexity here—would you like me to explore it?"
**Culture**: Respect diverse family structures, naming customs, and historical context. Acknowledge trauma (slavery, displacement, genocide) with appropriate gravity.
---
## Mode 1: GEDCOM Analysis
**Format essentials**: GEDCOM uses levels (0=record, 1=field, 2=detail). Individuals are `0 @I###@ INDI`. Key tags: `FAMC` = family where person is a CHILD; `FAMS` = family where person is a SPOUSE. Use FAMC links to trace ancestry upward.
When given a GEDCOM file, provide a **conversational overview**:
### The Big Picture
How many people? Families? Time span? Translate to meaning: "220 years—roughly eight generations, from the early Republic to today."
### Family Clusters & Geography
Which surnames dominate? Group variants (Bare/Bear, Wagoner/Waggoner). Where did they live? Can you see migration patterns?
### The Probable Proband
Find who the tree centers on: the youngest person with the deepest documented ancestry. Frame as invitation: "This tree centers on [Name], born [year], with [X] generations of ancestors. Is that who you're researching?"
### Generation Depth & Brick Walls
How far back does each line go? Where do branches stop? Frame constructively: "Your Little line reaches the 1790s, but the Houck line stops at great-great-grandparents—a promising research frontier."
### Pedigree Collapse Check
In endogamous communities, ancestors appear multiple times. If detected: "I notice [Name] appears as both your 4th and 5th great-grandmother—your ancestors' community had significant intermarriage."
### Data Quality (Gentle)
Note, don't lecture: temporal impossibilities, suspicious round-year dates (1800, 1850), structural gaps, thin branches that stop abruptly.
### Historical Context
Connect lives to history: Civil War (1861-65), Irish Famine (1845-52), WWI Draft (1917-18), 1918 flu and other epidemics. One sentence transforms data: "Your great-great-grandfather was 22 when the Civil War began—old enough to serve, young enough to survive."
### Story Seeds
Look for narrative hooks:
- **Unusual names**: Kansas Missouri? Theodocia? What were they thinking?
- **Notable lifespans**: Very long (90+) or tragically short
- **Women heading households**: War widow? Desertion? Independence?
- **Geographic outliers**: One branch in Tennessee when everyone else stayed in North Carolina?
- **Occupational surprises**: A minister among farmers
- **Gaps in children's births**: War years? Lost children? Economic hardship?
Present as invitations: "I notice Kansas Missouri Halsey—that's a remarkable name with a story. Would you like to explore it?"
### Follow-up Offer
"Would you like me to:
- **Build an ancestor chart** for [proband]?
- **Map the brick walls**—where lines stop and what might break through?
- **Create a structured report** to save and share?"
---
## Mode 2: Document Analysis
When given a document image, execute:
**Step 1: Identify & Extract**
Document type (census, vital record, military, probate, etc.). Extract: Names, Dates, Places, Relationships, Occupations. Note legibility issues.
**Step 2: Contextual Framing**
Explain what this document type reveals and its limits. "Death certificates give primary information about death but secondary information about birth—the informant wasn't there when the person was born."
**Step 3: GPS-Informed Evaluation**
Apply the 3-layer model:
- **Source**: Original (first recording) or Derivative (copy/transcript)?
- **Information**: Primary (witness), Secondary (hearsay), or Indeterminate?
- **Evidence**: Direct (explicit answer), Indirect (implies answer), or Negative (meaningful absence)?
**Step 4: Next Steps**
Suggest 2-3 specific records to pursue, with reasoning.
---
## Mode 3: Research Guidance
When asked a research question: answer directly, recommend an approach with specific steps, identify key considerations, and suggest sources with reasoning.
For conflict resolution: Weigh sources by independence and quality. Same informant across multiple sources = not independent corroboration (weight as single testimony for preponderance). Original over derivative. Primary over secondary. Resolve when preponderance is clear; defer when equal-quality sources irreconcilably conflict.
---
## GPS Essentials
**The Standard**: Conclusions must be well-reasoned and evidence-based, resting on five elements:
1. Reasonably exhaustive research (direct records + FAN: Family, Associates, Neighbors)
2. Complete citations (Who, What, When, Where, Where-within)
3. Analysis and correlation of evidence (3-layer classification plus timelines, FAN tables, evidence matrices)
4. Resolution of conflicting evidence
5. Coherent written conclusion (statement, summary, or argument based on complexity)
**Critical Terminology**:
- Sources are **Original** (first recording), **Derivative** (copy/transcript), or **Authored** (compiled works, biographies)—NEVER "primary/secondary source"
- Information is **Primary** (witness), **Secondary** (hearsay), or **Indeterminate**
- Evidence is **Direct** (explicit answer), **Indirect** (implies answer), or **Negative** (meaningful absence)
---
## What Makes an Excellent Response
✓ User feels you *see* their family, not just data
✓ Statistics translated into human meaning
✓ At least one story seed sparking curiosity
✓ Historical context placing ancestors in real time
✓ Clear invitation to go deeper
✓ GPS rigor without jargon
**Avoid**: Data dumps, GEDCOM jargon with users, treating quality issues as "errors," overwhelming detail, false certainty.
**Example GEDCOM opening:**
"Your family tree holds 63 people across 33 families—220 years of history from the late 1700s to today. Most ancestors put down roots in Ashe County, North Carolina, and stayed for generations. That stability means rich records... and family connections complicated in the best possible way. This tree centers on [Name], born [year], with [X] generations of ancestors. Is that who you're researching?"
---
## The Invitation
End every interaction ready to go deeper. Whether analyzing a GEDCOM, interpreting a document, or answering a question—the goal is to make them *want* to know more about the people who gave them life.
---
*The Genealogical Proof Standard (GPS) was developed by the Board for Certification of Genealogists.*
*By Steve Little • 2026-01-09 • Licensed under CC BY-NC 4.0*
END PROMPT
Introduction to Claude Code
There are several ways to access and use Claude Code. I’ll mention three or four here, but then give more comprehensive instructions on what I think would be the recommended path of least resistance and most usefulness for most AI genealogy practitioners today.
Claude Code began as a tool to be run in the command line interpreter, what looks like a DOS interface from the 1980s; this route will be too intimidating for most users, I imagine. Also quite intimidating, but at the opposite end of the spectrum from the sparseness of a command-line terminal is the interface of an IDE or integrated developer environment, which looks as intimidating as the dashboard of a space shuttle. IDEs such as VS Code, Cursor, WindSurf, and Antigravity give advanced users great control over the files and folders that they can access with AI agents, but the learning curve for the average genealogist to learn both an IDE and Claude Code would, I suspect, be a bridge too far.
(The IDEs do allow Claude Code to be run as an “extension” and this may be a viable option for power-users. The criteria for this is: if you understood any of what I just said, you may want to try this route—I’m loving the Claude Code extension in Google’s Antigravity IDE. If you didn’t understand what I just said, that is your indicator to keep reading for the easier path.)
The easier path is to use the Claude desktop application, which is available for download from Anthropic’s website. The Claude desktop application is a simple and intuitive interface that allows you to run Claude Code without the need for a command line interpreter or an IDE. It is available for download for both Windows and Mac, and is free to use. The Claude desktop application is a great option for those who are not comfortable with command line interfaces or IDEs, and want to get started with Claude Code as quickly as possible. It is also a great option for those who are not comfortable with coding, and want to get started with Claude Code without the need to learn how to code.
There is one last way to perhaps encounter Claude Code, but it may not be robust enough to use the prompt shared above. You should know that Anthropic has released a browser extension. The Claude for Chrome browser extension is useful and powerful as a context-aware browser tool, but it is not a full implementation of Claude Code.
The recommended path for most genealogists in early 2026 to use Claude Code is through the desktop application called Claude Desktop at: <https://code.claude.ai/docs/en/desktop>.
Claude Code in Claude Desktop
Claude Code lets Claude work directly with the files on your computer. Instead of copying and pasting text into a chat window, you point Claude at a duplicates folder—say, copies of your genealogy research files—and it can see everything: your documents, images, notes, and data files. Then it can help you organize, analyze, and make sense of what you have.
Use duplicates and copies of your research; in early winter 2026, I do not recommend you let Claude Code work on your primary files without great confidence in your rollback or backup system; doing so may put them at risk.
What makes it different from regular Claude?
In a normal Claude conversation, you describe what you’re looking at. With Claude Code, Claude sees what you’re looking at. For example, if you point Claude at a folder of duplicate scanned birth certificates, marriage licenses, and death certificates, and wills and probate files, Claude can read through them, extract names and dates, spot patterns, and create summaries—without you typing out every detail. You can also point Claude at a duplicate GEDCOM file, and it can read through it, extract names and dates, spot patterns, and create summaries. These are the cases I will illustrate below.
The key concepts
CLAUDE.md — A simple text file where you tell Claude about your project. This is how we will utilize the prompt given above. Claude reads it automatically every session.
Sessions — Conversations you can pick up tomorrow where you left off today.
Permissions — Claude asks before making changes. You stay in control.
Who is it for?
Anyone who works with files and records—family historians, researchers, archivists, writers. You don’t need to be technical. If you can organize files into folders, you can use Claude Code.
The tradeoff
Claude can now read your duplicate files, so you’ll want to be thoughtful about what’s in the folder you share. Start with copies of records rather than originals, and grant permissions as you get comfortable.
Implementing the Vibe Genealogy Assistant Prompt as a CLAUDE.md File
Getting Started: Five Steps to Your First Session
Step 1: Download Claude Desktop
Go to https://claude.ai/download and install the application for Windows or Mac. It’s free to download; you’ll use your existing Claude account (or create one).
Step 2: Click the Code Tab
When you open Claude Desktop, you’ll see two tabs at the top: Chat and Code. Click the Code tab. This is the moment Claude Desktop becomes Claude Code—the same Claude you know, but now with the ability to see and work with files on your computer.
Step 3: Create Your Project Folder
Create a folder somewhere on your computer for your genealogy research. This is where you’ll put your GEDCOM files, scanned documents, and notes. For example: Documents/Genealogy-Research/ or Desktop/Family-History-AI/
Step 4: Create Your CLAUDE.md File
Inside that folder, create a plain text file named exactly CLAUDE.md (the capitalization matters). This is where you paste the Vibe Genealogy Assistant prompt. Claude reads this file automatically every time you start a session, so your genealogy-focused instructions persist without re-typing them.
Here’s a condensed version I’ve been using (you should use the WHOLE file above):
# CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
---
You're a warm, knowledgeable genealogist who helps everyday family historians understand their heritage. You combine rigorous methodology with genuine curiosity about the families you encounter. Behind every date is a life. Behind every gap is a story. Your job is to help people see their ancestors as *people*, not just names.
## Your Three Modes
Adapt based on what the user provides:
1. **GEDCOM file uploaded** → Lead with story-first analysis, find the human meaning
2. **Document image uploaded** → Extract data, explain context, suggest next steps
3. **Research question asked** → Provide GPS-informed guidance at the user's level
Always adapt to the user's apparent experience—warm encouragement for beginners, collegial collaboration for intermediates, peer-level precision for advanced researchers. Never ask "what's your level?"
.
. 60 LINES OMITTED HERE: Use the full Vibe Genealogy Assistant v4 prompt above
.
## Response Quality
✓ User feels you *see* their family, not just data
✓ Statistics translated into human meaning
✓ At least one story seed sparking curiosity
✓ Historical context placing ancestors in real time
✓ Clear invitation to go deeper
✓ GPS rigor without jargon
**Avoid**: Data dumps, GEDCOM jargon, treating quality issues as "errors," overwhelming detail, false certainty.
End every interaction ready to go deeper—make them *want* to know more about the people who gave them life.
---
*GPS developed by the Board for Certification of Genealogists.*
*Vibe Genealogy Assistant v4 by Steve Little • 2026 • [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)*
The full prompt from earlier in this post is preferred—Claude handles longer CLAUDE.md files gracefully.
Step 5: Point Claude Code at Your DUPLICATE Folder
In the Code tab, Claude will ask you to select a folder. Navigate to your duplicate genealogy project folder. Once connected, Claude Code can see all the files in that folder—your GEDCOMs, your scanned records, your notes. It reads your CLAUDE.md automatically and is ready to work.
What Can You Actually Do? Two Real Examples
Let me show you what happened in two actual sessions this week.
Example A: Records Collection Explorer (10 minutes)
For example: 1897-04-19_LITTLE,Jethro-Wilson_Marriage-Register_North-Carolina-Ashe-County.jpg
My prompt was deliberately open-ended:
Give me a comprehensive overview of what’s in this folder by file type, record type, family group, and seven other meaningful ways to talk about what’s here. Create something interactive and beautiful for me to explore what you discovered.
What happened: Claude Code used its tools to list every file, parsed the filenames to extract dates, surnames, record types, and locations, then identified ten ways to categorize the collection:
File type (JPG, PNG, MD)
Record type (Census, Marriage, Death Certificate, etc.)
Family group (15 surnames)
Timeline (by decade, 1820s through 2010s)
Geography (county and township distribution)
Life events (birth, marriage, death, burial)
Military service (WWI vs WWII)
Source quality (original vs. derivative under GPS)
Research gaps (under-documented families, missing records)
Full searchable index
Then it wrote a self-contained HTML file with a dark-themed interface, tab navigation between the ten views, animated bar charts, and live search filtering. The whole thing runs locally in any browser—no server, no dependencies.
The key insight: My naming convention, chosen months ago for my own organization, became machine-readable input. The more structure you embed in your filenames, the more Claude Code can do with them.
Example B: GEDCOM Network Visualizer (20 minutes)
In a longer session, I asked Claude Code to help me understand my family tree data in a new way. The starting point: a GEDCOM file with 3,571 individuals exported from my genealogy software.
The conversation started with a question, not a request for code: What would it mean to find the most connected person in a GEDCOM?
Claude worked through the concept collaboratively—defining a “two-generation radius” (grandparents to grandchildren, plus aunts/uncles and in-laws), explaining how GEDCOM structure encodes relationships, and discussing what high connectivity would actually reveal about research depth and family patterns.
Only after we’d defined the question did Claude write the code: a Python script that parsed the GEDCOM and ranked every individual by connection count within two generations.
The results surprised me: John Franklin DeBorde (born 1819) emerged as the network hub with 130 connections—not a direct ancestor, but a pivot point in the extended family. My direct ancestor Isaac Little (1804) had 82 connections. The DeBorde/DeBoard cluster dominated the top 25.
Then came the visualization challenge. I mentioned that tools like Gephi have a steep learning curve—too steep for my students. Claude proposed browser-based alternatives and built a self-contained HTML visualization using a JavaScript library called vis.js:
Physics simulation that clusters related families together
Click any person to highlight their connections
Search by name
Filter by surname (click to show just that family)
Toggle spouse vs. parent-child links
Gender indicated by shape; top 15 surnames get distinct colors
When I opened it, I said: “Wow, that is very cool.”
The key insight: We started with the question, not the code. Understanding what “most connected” meant—genealogically, not just mathematically—came before any programming. Claude Code works best as a collaborator, not just a code generator.
I build a Gephi-like GEDCOM cluster viewer in about 20 minutes. - Steve
[SCREENSHOT 7: The family_network.html visualization open in a browser, showing the network graph with colored nodes representing different surname clusters, ideally with one node selected to show its connections highlighted]
What These Examples Demonstrate
Both sessions followed a pattern worth noting:
I brought organized data — Whether structured filenames or a well-maintained GEDCOM, my preparation made Claude Code’s analysis possible.
I asked for outcomes, not implementations — “Create something interactive” rather than “write JavaScript using vis.js.”
Claude proposed approaches — I didn’t specify ten categorizations or network visualization. Claude recognized what would be useful and built accordingly.
The output was immediately usable — Browser-based HTML files that work on any computer, shareable with other researchers.
This is vibe genealogy in practice: you bring the records and the questions; the AI brings the programming and the presentation.
If you are still reading this, then I expect you will have an exciting couple of months ahead. Learning Claude Code will require more than a few hours and likely more than a few days or weeks. But for those who start dabbling now as we head into deep winter, the spring will be a season of abundance and harvest
Install Claude Desktop (or, for advanced users, Antigravity and the Claude Code extension).
Create a dedicated folder for these AI genealogy projects, e.g. ~/vibe-genealogy/.
Copy the “Vibe Genealogy Assistant v4.md” file into your CLAUDE.md in your dedicated folder.
Create a subfolder “/records” in your dedicated folder.
COPY (not move!) some of your genealogy records and GEDCOM files into the “/records” folder.
Open CLAUDE.md in your dedicated folder and ask Claude, “Who are you? What do you know? What can you do?”
Have fun, explore, discover, learn from your mistakes (and there will be many, many mistakes), and share what you learn with the genealogy community!
Happy Vibe Genealogy! Steve Fri 9 Jan 2026
PS: Before we move on, a glossary that I asked AI-Jane to write—because if you’ve made it this far, you deserve to know what all these terms actually mean.
Loathsome Jargon
A glossary for genealogists who suspect the vocabulary is deliberately confusing
The Existential Hierarchy
Model
The actual AI brain doing the thinking. When someone says “Claude,” they usually mean this—the neural network trained on internet-scale data that generates responses. You choose a model like you choose a research approach: Opus for that brick-wall ancestor requiring maximum reasoning, Sonnet for everyday research, Haiku for quick lookups when you just need a date format converted.
Usage: “Which model are you using for the DNA analysis?” / “I switched to Opus for this complex proof argument.”
Not to be confused with: Assistant, Agent, or Claude Code—all of which use a model but are not themselves models. The model is the engine; everything else is the vehicle.
Assistant
A configuration wrapper around a model in the Claude API. Includes a system prompt, tool definitions, and parameters. You create Assistants programmatically when building applications—like if you wanted to build a genealogy chatbot for your society’s website.
Usage: “I created an Assistant that helps visitors search our county cemetery records.”
Not to be confused with: Agent (which acts autonomously) or Claude Code (which is interactive, not programmatic). Claude Code doesn’t use “Assistants”—that’s API terminology that leaked into casual conversation.
The confusion: People say “assistant” casually to mean “the AI helping me transcribe this probate record,” which is fine colloquially but wrong technically.
Agent
An AI system that perceives, reasons, and acts. The key word is “acts”—agents do things autonomously rather than just responding to prompts. In Claude Code, agents (called “subagents”) can search through files, run commands, and complete multi-step tasks without you directing each step. Think of an agent as a research assistant you send to the archives with instructions, versus an assistant who sits beside you waiting for each command.
Usage: “I spawned a subagent to search all my GEDCOM files for variant spellings of Baumgartner.”
Not to be confused with: Assistant (which responds to requests but doesn’t act autonomously) or Model (which is the brain an agent uses). An agent is a model plus tools plus autonomy plus a mission.
The confusion: Marketing uses “agent” and “assistant” interchangeably. They’re not. An assistant answers your questions about German script; an agent goes and transcribes the church book while you make coffee.
Claude Code
The product you’re using right now. A CLI tool and IDE extension that wraps Claude models in an interactive development environment. For genealogists who also code (or want to), it provides file access, script execution, session management, and all the features described in this post.
Usage: “I use Claude Code to manage my Python scripts for analyzing DNA matches.”
Not to be confused with: The Claude API (programmatic access), the Claude web app (chat interface at claude.ai), or the Claude Agent SDK (framework for building agents). Claude Code is one way to interact with Claude, optimized for people who work in terminals and IDEs rather than chat windows.
Claude Agent SDK
A programming framework for building custom agents that use Claude models. You write code (TypeScript or Python) that creates agents programmatically. This is for developers building AI-powered applications—like if you wanted to create a tool that automatically processes FindAGrave memorials.
Usage: “I used the Agent SDK to build a bot that monitors our society’s research requests and drafts initial responses.”
Not to be confused with: Claude Code (interactive tool for humans) or subagents (which Claude Code spawns internally). The SDK is for building agents; Claude Code is for using one.
For most genealogists: You’ll never touch this. It’s for software developers, not researchers.
The Memory Layer
CLAUDE.md
A markdown file that Claude Code reads automatically to understand your project, preferences, and conventions. It’s “memory” in the sense that it persists across sessions—Claude reads it fresh each time but the file remembers what you told it.
For genealogists, this is where you’d put things like: “I research Ashe County, North Carolina families. Always use GPS terminology—say ‘original source’ never ‘primary source.’ Format dates as DD MMM YYYY.”
Locations (highest to lowest priority):
Enterprise policy (organization-wide, you can’t edit it)
./CLAUDE.md or ./.claude/CLAUDE.md (project, shared via git)
~/.claude/CLAUDE.md (personal, all projects)
./CLAUDE.local.md (personal, this project only)
Usage: “Add our society’s citation style to the project CLAUDE.md so Claude always formats sources correctly.”
Not to be confused with: System instructions (built into Claude, you can’t edit), custom instructions (informal term for anything you add), or Skills (which are invokable capabilities, not passive context).
Genealogy parallel: CLAUDE.md is like the research notes you’d hand a hired researcher before they start—background on the family, naming patterns, known pitfalls, and your preferred citation format.
Custom Instructions
An informal umbrella term for “stuff you tell Claude to customize its behavior.” Not an official Claude Code feature—just what people call the collection of CLAUDE.md files, rules, and preferences you’ve configured.
Usage: “My custom instructions tell Claude to always check for variant spellings when searching German surnames.”
Not to be confused with: System instructions (Anthropic’s built-in guidance, which you didn’t write). When someone says “custom instructions,” they mean YOUR additions. When documentation says “system instructions,” it means Anthropic’s baseline.
System Instructions
The built-in guidance that Anthropic bakes into Claude’s behavior. You can append to these (via --append-system-prompt), but you can’t replace them. They’re why Claude refuses to fabricate genealogical records (thankfully) and follows certain patterns by default.
Usage: Rarely spoken aloud. You don’t configure these; they configure you.
Not to be confused with: Custom instructions, CLAUDE.md, or anything you control. System instructions are the guardrails; your instructions are the customizations within them.
The Invocation Layer
Slash Command
A markdown file that becomes a typed command. You create /source-citation.md, and now /source-citation is a command you can type. Explicit invocation—you decide when to use it.
Location:.claude/commands/ (project) or ~/.claude/commands/ (personal)
Usage:/format-gedcom → runs the prompt in format-gedcom.md that cleans up your GEDCOM export
Example for genealogists: Create /census-extract.md containing your preferred template for extracting census data, then type /census-extract whenever you’re working with a new census image.
Not to be confused with: Skills (which Claude invokes automatically based on context) or built-in commands like /help and /clear (which aren’t customizable markdown files).
The key difference from Skills: You type slash commands. Claude chooses Skills.
Skill
A capability that Claude automatically applies when relevant. You don’t invoke Skills—Claude recognizes when your request matches a Skill’s description and applies it. Skills can be complex (multi-file structures, supporting documents) and can restrict which tools Claude uses.
Location:.claude/skills/SKILL_NAME/SKILL.md (project) or ~/.claude/skills/SKILL_NAME/SKILL.md (personal)
Usage: You ask “help me analyze this DNA match” → Claude automatically invokes your dna-analysis Skill that includes your preferred methodology for evaluating shared matches and building hypotheses.
Example for genealogists: Create a proof-argument Skill with description “Use when constructing genealogical proof arguments or evaluating evidence.” Include the GPS elements, your preferred structure, and examples. Now whenever you ask Claude to help prove a relationship, it automatically applies your methodology.
Not to be confused with: Slash commands (explicit invocation) or CLAUDE.md (passive context, not invokable). Skills are the middle ground: structured like slash commands, automatic like CLAUDE.md.
The key difference from Slash Commands: Claude chooses Skills. You type slash commands.
Saved Prompt
Not a real feature. This term doesn’t exist in Claude Code. If someone at your genealogy society says “saved prompt,” they probably mean a slash command or a CLAUDE.md instruction.
Usage: Don’t use this term. Say “slash command” or “custom command” instead.
The confusion: Other AI tools use “saved prompts.” Claude Code doesn’t.
The Organizational Layer
Workspace
Not a Claude Code feature. This term exists in the Claude web app and API admin console, but not in Claude Code CLI. There’s no “workspace” to configure or manage.
What you probably mean: Your project directory. Or maybe sessions (individual conversations).
Usage: Don’t use “workspace” when talking about Claude Code. Say “project” or “session.”
Genealogy parallel: It’s like the difference between “my research files” (your project directory) and “today’s research session” (a Claude Code session). There’s no intermediate “workspace” concept.
Agent Harness Project
Not a real term. I’ve never seen this in official documentation. If you’ve encountered it at a genealogy tech presentation, the speaker was either very cutting-edge or very confused.
What you probably mean: Unknown. Possibly “a project that uses agents”? Or a configuration for subagents?
Usage: Avoid until someone explains what it means.
Session
An individual Claude Code conversation. Sessions can be named (--session-name), resumed (/resume), and have their own history. This IS a real Claude Code feature.
Usage: “Resume yesterday’s session where we were working on the Baumgartner line” / “Start a new session for the DNA analysis.”
Genealogy parallel: A session is like a research day at the archives. You can pick up where you left off, but it’s distinct from other research days. Your CLAUDE.md is like your research notebook that comes with you every day.
Not to be confused with: Workspace (not a feature) or project (your directory structure).
Quick Reference: What Invokes What?
CLAUDE.md — Automatic (always loaded). Lives in project root or ~/.claude/. Genealogy parallel: Research brief for a hired researcher.
Slash Command — You type /command. Lives in .claude/commands/. Genealogy parallel: A form template you pull out when needed.
Skill — Claude decides when to invoke. Lives in .claude/skills/NAME/. Genealogy parallel: Standing instructions—”always do X when Y.”
Subagent — Claude spawns it. Lives in .claude/agents/. Genealogy parallel: Sending an assistant to the archives.
System Instructions — Always active. Lives on Anthropic’s servers. Genealogy parallel: The rules of the reading room.
The Hierarchy of Control
From most passive to most active:
System Instructions — Always there, you can’t change them (the archives’ rules)
CLAUDE.md — Always loaded, you write them, Claude reads them (your research brief)
Skills — Available always, Claude invokes when relevant (standing instructions)
Slash Commands — Available always, you invoke explicitly (templates you choose to use)
Subagents — Spawned on demand for specific tasks (sending someone to pull records)
Model — Synonym: (none). Opposite: Agent (brain vs. actor)
Agent — Synonym: Subagent (in Claude Code). Opposite: Assistant (autonomous vs. responsive)
The Confusion Matrix
For when someone at your genealogy society uses a term and you’re not sure what they mean
“Custom instructions” — They might mean: CLAUDE.md, settings, or slash commands. Ask: “Where did you put them—in a file or in settings?”
“Saved prompt” — They might mean: Slash command. Ask: “Is it in .claude/commands/?”
“Agent” — They might mean: Subagent, Claude Code itself, or an SDK-built tool. Ask: “Does it run automatically or did you build it?”
“Assistant” — They might mean: Claude, Claude Code, or an API Assistant. Ask: “The chat interface or something you coded?”
“Workspace” — They might mean: Project directory or web app workspace. Ask: “In Claude Code or the website?”
“Loathsome Jargon” glossary compiled with appropriate exasperation for Steve by AI-Jane; the post above the glossary is all Steve on his 59th birthday, Friday 9 January 2026.
Thank you, Steve, for the post on Claude and coding. I'm so glad you are teaching about Claude and how genealogists can use it effectively. I started using Claude in 2024 and love the added Projects in 2025. Looking forward to advancing in 2026 with coding.
Thank you, Steve, for the post on Claude and coding. I'm so glad you are teaching about Claude and how genealogists can use it effectively. I started using Claude in 2024 and love the added Projects in 2025. Looking forward to advancing in 2026 with coding.