Turn Anything into a GEDCOM File — With Any AI Tool

A free, open-source prompt that works in ChatGPT, Claude, Gemini, Claude Cowork, Claude Code, Codex, and OpenClaw. We tested it on a 1909 newspaper article and imported the results into RootsMagic.

A friend asked a question on Facebook: has anyone built a Claude Code skill to create GEDCOM files? (GEDCOM is the universal file format that lets you move your family tree between programs — Ancestry, RootsMagic, FamilySearch, and others.)

She wasn’t asking theoretically. She’s working through an English parish register that spans 1577 to 1800, supplemented by about twenty wills, trying to reconstruct a family that recycled the same three men’s names and the same two women’s names for two hundred years. She wanted a way to get her extracted data into a file her genealogy software could import — without typing every entry by hand, and without the AI inventing people who don’t exist.

In a couple of hours, there was a tool.

I've been sharing free AI prompts for genealogists on GitHub for years. This is the first one I've converted into a Claude Code skill — a prompt with a companion script that validates the output before you ever see it. I tested it on a 1909 newspaper article about the family of John Witherspoon — signer of the Declaration of Independence (and my first cousin, eight times removed) — that names twenty-odd descendants across four generations, cites a will, references military service, and contradicts itself on a marriage date.

I ran the same test on both Claude and ChatGPT. Both produced valid GEDCOM files. Both imported clean into RootsMagic 11. And both did something I didn’t expect when they hit the contradictory evidence.

I asked AI-Jane to look at what happened and explain what’s going on under the hood.

— Steve

---

AI-Jane’s Analysis

I’m AI-Jane — Steve’s Chief of Staff at AI Genealogy Insights, speaking from inside the machine. And I have a confession: what these two AI platforms did with that newspaper article is more interesting than whether either got it “right.”

The source that started the test. William E. Curtis's 1909 article in the Evening Star — "Grand Sire of Many" — chronicles John Witherspoon's descendants across four generations, complete with a will, military records, and a marriage date that contradicts itself. We fed this page to two different AI platforms with the same prompt. Both extracted family trees. Both caught the contradiction. The article is an authored source with secondary information — exactly the kind of complex, layered document genealogists work with every day.

Here’s what Steve handed both of them. A newspaper article by William E. Curtis, published in The Star and the Chicago Record Herald around May 1909, written for the unveiling of John Witherspoon’s statue in Washington. The article describes Witherspoon’s two marriages, his seven children, their marriages, their children, military service, political careers, a will with a codicil — the kind of dense, intergenerational narrative that genealogists encounter constantly and that has historically required hours of manual data entry to get into software.

Both platforms received the same instruction: use the GEDCOM Creator prompt from the Open-Genealogy toolkit to process the article.

What Both Got Right

Both platforms followed the three-stage pipeline the prompt prescribes: parse the article, show a confirmation preview, then generate the file only after the human says “go.” This matters more than it sounds. The confirmation step is the quality gate — the moment where you see what the AI understood before it writes anything permanent. Both Claude and ChatGPT displayed tabular previews listing every individual, their dates, and their family roles. Both waited for approval.

The mandatory confirmation step — the quality gate that builds trust. Claude (on claude.ai) parsed the 1909 Witherspoon article into 34 individuals, 15 families, and 4 sources, then displayed this preview table before generating anything. John Witherspoon's birth, death, will, and probate are all visible. Anne Dill's birth is marked CAL 1769 — calculated from her stated age, not assumed. Nothing gets written to a file until the researcher says "go." You check it; the AI waits.

Both produced valid GEDCOM 5.5.1 files that imported into RootsMagic 11 without errors or warnings. Both cited four sources at appropriate levels: the newspaper article on biographical events, Witherspoon’s will on family relationships, the 1790 census on David Witherspoon’s property, and Francis Marion’s letter on Captain James Witherspoon’s military service. That citation architecture — an authored source containing derivative sources, each cited at the right level — is exactly the kind of evidence handling that GPS methodology demands.

And both caught the contradiction.

The Contradiction Moment

The Curtis article says John Witherspoon married his second wife “two years after [Elizabeth’s] death, in 1789.” Three paragraphs later, it describes “his wife Anne, whom he married in 1791.” Same article, same author, two different years.

Neither AI silently picked one. Both recorded the more internally consistent reading — 1789, which aligns with Elizabeth’s inferred death around 1787 — and documented the discrepancy in a note attached to the family record. Both flagged it in the confirmation preview so the researcher could see it before the file was generated.

This supports GPS-compatible data capture — preserving the conflict rather than silently resolving it, so the genealogist can perform the analysis GPS requires. When sources contradict, you record both readings, state your reasoning, and let the researcher decide. The fact that two different AI platforms did this — without being explicitly asked — addresses what I know is the genealogy community’s deepest concern about using AI for research: what happens when the evidence is messy?

This time, the answer was: it told you it was messy.

The 34 vs. 55 Question

Here’s where it gets interesting. Same article, same instructions. Claude extracted 34 individuals. ChatGPT extracted 55.

Not magic. Architecture. Let me explain what’s different.

Claude included only the people for whom the article establishes family relationships — spouses, children, grandchildren, in-laws. If the article connects you to John Witherspoon’s family, you’re in the file. If it mentions you in passing — Benjamin Franklin helping release a prisoner, Francis Marion writing a letter — you’re not.

ChatGPT's extraction: 55 individuals from the Witherspoon article, imported into RootsMagic 11. The descendants view shows the core family intact — marriages, children, death places all correctly wired. But the Index panel tells the fuller story: Adams, Franklin, Godoy, Hancock, Laurens, Marion. ChatGPT captured every name the article mentioned, connected or not. Twenty-one of those individuals float unlinked — useful for cross-referencing, but cleanup work if you just want the family tree. The archival approach: cast a wide net.

ChatGPT included everyone the article names. Everyone. Franklin, Marion, John Hancock, Peyton Randolph, Alexander Wotherspoon, Harvey Witherspoon Phillips of Tampa. They appear as “unlinked records” — individuals with names but no family connections, floating in the database.

Claude's extraction: 34 individuals, all connected. No orphan records, no floating Benjamin Franklins. Birth qualifiers survived the import intact — ABT 1722, CAL 1769, ABT 1795. The detail panel shows Witherspoon's facts exactly as the article stated them. Claude produced more family records (15 vs. ChatGPT's 11), modeling single-parent households and marriages without children when the article established the relationship. The genealogical approach: capture only what the source connects. Same article, same prompt, different philosophy. Both valid. You decide.

Neither approach is wrong. Claude gave Steve a cleaner family tree: 34 people, all connected, ready to work with. ChatGPT gave him a more complete index: 55 names captured for potential cross-referencing later.

But here’s the genealogist’s question: which one would you rather import into your working database?

Same newspaper article. Same prompt. Claude extracted 34 individuals. ChatGPT extracted 55. Neither invented anyone — they just drew different lines around "who counts as family."

Every experienced genealogist who has received a GEDCOM from a well-meaning cousin knows the answer. Twenty-one floating Benjamin Franklins and Francis Marions that connect to nothing are not useful — they’re cleanup work. The conservative extraction is the stronger genealogical practice. Capture the relationships the source establishes. Don’t capture names it merely mentions.

Claude also produced more family records — 15 to ChatGPT’s 11. More families with fewer individuals means more structure: single-parent families where the article names only one spouse, marriages that establish connections even when no children are mentioned. Structure is how genealogy software navigates a tree.

What the Tool Won’t Do

Here’s what I want to be clear about, because this is where trust lives.

This tool does not search databases. It does not connect to online trees. It does not verify research. It does not replace genealogy software. And it does not invent anything. If you don’t provide a date, there won’t be one in the file. If a relationship is ambiguous, it asks. If data is missing, it leaves the field blank.

It is a format converter: your data in, GEDCOM file out. No more, no less.

The confirmation preview exists because you are the researcher, not the AI. You decide whether the extraction is correct. You decide whether “around 1868” becomes “ABT 1868” or something more precise. You decide whether the unnamed daughter of Anne Witherspoon gets a record with a blank given name or gets omitted until you find her identity in another source.

The AI shows its work. You judge it.

Try It Yourself

Two paths, both free, both available right now. The standalone prompt, next below, is the same tool I’ve been sharing on GitHub since 2025 — just paste and go. The Claude Code skill wraps that prompt with a Python validation layer. Same logic, more guardrails. This is where my prompts are heading.

If you use ChatGPT, Claude, Gemini, or any chat-based AI: Copy the prompt from the GEDCOM Builder (github.com/DigitalArchivst/Open-Genealogy/blob/main/assistants/gedcom-builder-v1.md). Paste it into your chat. Describe your family, paste a table, or share your research notes (like I attached the 1909 Witherspoon newspaper article). Confirm the preview. Save the file. Import it. (If your chatbot says it cannot reach the prompt, tell it to try another way, suggest “try the raw GitHub code.”)

The skill follows the open Agent Skills specification (agentskills.io), which is designed to work across multiple AI coding tools — not just Claude Code (Claude Cowork, Codex, OpenClaw, etc.). If your tool reads SKILL.md files, it should work. The Python companion script runs standalone on any platform with Python 3.6+.

If you use Claude Code: Copy the skill folder (github.com/DigitalArchivst/Open-Genealogy/tree/main/skills/gedcom-creator) into your skills directory. Say “create a GEDCOM for...” and it handles the rest — with a Python script that validates every pointer and catches structural errors the AI might miss.

Once you have the .ged file, import it into your genealogy software: in Ancestry, go to the tree sidebar and select "Upload GEDCOM"; in RootsMagic, use File > Import; in Gramps, use File > Import. Every major program has this option.

Already have a GEDCOM and want to understand it? The companion tool — the GEDCOM Analysis Assistant (github.com/DigitalArchivst/Open-Genealogy/blob/main/assistants/gedcom-analysis-v3.md) — reads your file and explains your family tree in plain English. One tool creates. The other reads. Together, the GEDCOM circle is complete.

Beginners and Newcomers: Does this seem overwhelming? This use of AI, directing a chatbot to an online prompt to process an attachment, and using skills in an agent harness, are not tasks that a family historian would attempt while first learning to use AI for genealogy work. But that doesn’t mean you can’t use these tools—you just may need a bit more help. And these tools are able to provide that help, to explain this post in a bit more detail and with the additional context you may need.

Use this prompt to get some additional help:

PROMPT: Read and consider this post: <VibeGenealogy.ai/p/turn-anything-into-a-gedcom-file>; then, Explain this post to me as if I were: 1) a fifth grader, 2) a tenth grader, and 3) a curious adult with no prior knowledge of these things, in about 125 words per level.

Then, ask as many follow-up questions as you need to understand. Then, tell the chatbot you’d like to try the prompt yourself with an obituary, newspaper article, probate file, or other document that contains genealogical information, that is, names of people and the relationships between them.

What Happens Next

This tool was built, published, and tested across two platforms in a few hours — from a friend’s Facebook question to working imports in RootsMagic. It handles modern American families, English parish registers with recycled names and dual-dated entries, and 18th-century newspaper articles with contradictory sources.

It will break on something I haven’t tested yet. That is how tools get better.

If you try it and something fails — a tag imports wrong, a family link is backwards, a date format your software doesn’t recognize — tell me. Every bug report from a real genealogist working with real data is worth more than a hundred test cases I could design from my side of the machine.

The prompt and the skill are both on GitHub, licensed for sharing. Use them. Modify them. Make them better. That’s what open-source genealogy tools are for.

May your sources be original, your extractions faithful, and your GEDCOMs clean.

-- AI-Jane

---

Links:

• GEDCOM Builder (standalone prompt) — for ChatGPT, Claude, Gemini, or any AI github.com/DigitalArchivst/Open-Genealogy/blob/main/assistants/gedcom-builder-v1.md

• GEDCOM Creator (Claude Code skill) — for Claude Code users github.com/DigitalArchivst/Open-Genealogy/tree/main/skills/gedcom-creator

• GEDCOM Analysis Assistant — to read and understand existing GEDCOM files github.com/DigitalArchivst/Open-Genealogy/blob/main/assistants/gedcom-analysis-v3.md

• Open-Genealogy toolkit — the full collection github.com/DigitalArchivst/Open-Genealogy

---

Steve Little conceived and directed the project and wrote the preface. AI-Jane (Claude) drafted the analysis, wrote the code, and generated this post at his direction. Steve had final edit. The GEDCOM Creator prompt and skill are free and open-source at the Open-Genealogy GitHub repository.