Lesson 3 of "How to Use AI for Family History"
In Lesson 2, we built a Town Profile and a Records Roadmap to learn the place and find where the records live. Now we start sifting the soil: you open a digitized document and find a wall of unstructured, messy text. It might be a dense census page, a sprawling old obituary, a will, a deed, or a parish register entry. Reading line by line, pulling out names, dates and relationships, and keeping your notes straight can make your eyes cross.
Bring from Lesson 2:
- Beginner: your checked short Worcester Town Profile, your Records Roadmap, and the map you opened, listed under "Map sources." In Lesson 3 you can add one sentence from the profile to your prompt.
- Intermediate: your checked detailed Lancaster Town Profile, your Records Roadmap, and your "Map sources" list, saved in your notes.
In this lesson, AI becomes your clue sifter. It does the sorting, and you do the thinking. By the end, you'll have turned a messy record into a clean table, and you'll know three new qualifiers that make the result easy to check.
Why Messy Records Wear You Out
When names, ages, and birthplaces are buried inside a paragraph or a raw transcription, it's easy to miss a vital clue: a servant, a visiting relative, a middle name, a neighbor who witnessed a document. Rebuilding that text into a table by hand is slow, and tired eyes make mistakes.
This is where your earlier work pays off. The Records Roadmap told you which records to look for. The Town Profile tells the AI what the place, period, and jurisdiction were, so it can make sense of unfamiliar place names and terms. When you paste a checked Town Profile in with your document, you're giving it the context a good research assistant would need.
Your Clue Sifter: Sorting, Not Interpreting
When you hand AI a messy document, you are not asking it to interpret the evidence or prove your family tree. You are asking it to do the sorting: pull out names, dates, roles, and relationships, and lay them in clean rows. That saves you transcription fatigue and lets your human brain get straight to the fun part, which is analyzing what the evidence means.
The research-intern rule from Lesson 1 still applies. An intern can build you a tidy table. An intern can also misread a smudged name and not tell you. So the table is a working copy, and the original document stays the authority, left exactly as you found it.
Before You Paste: Text First, Privacy Always
Three steps before your first extraction prompt:
- Start with typed text. The safest input is a typed transcription, either your own or the one a records website provides. If you upload a photo of handwriting instead, treat the AI's reading as a rough draft. It can misread letters, swap one name for a similar one, or fill in words it can't see. Try transcribing a few lines yourself first and compare.
- Protect living people. This is the Lesson 1 rule in action. Older census pages and wills are generally safe to paste. Recent obituaries often name living relatives, so replace those names with placeholders such as "[living daughter]" before you paste anything.
- Note where it came from. While the page is still open, write down the website or archive, the collection, the page or household number, and the date you looked. Lesson 4 shows how these details become a citation.
Also check which tools can accept images and how they handle your data. Those features differ by service and change often.
Level Up Your Prompt Formula: Three Extraction Qualifiers
The formula from Lesson 2 still holds:
ROLE + CONTEXT + TASK + FORMAT + GUARDRAIL (w/3 qualifiers)
= AN ORGANIZED TABLE
For extraction work, the guardrail gets three new qualifiers, and each one makes the AI's table easy to check:
- "As written." Ask for names and spellings exactly as they appear in the text. AI tends to quietly "fix" an old spelling, which can erase the very clue you need.
- "Exact words." Add a column that quotes the words from the text behind each row. If the quote isn't in your document, you've caught a mistake in seconds.
- "Not stated." Tell the AI to write "not stated" when the text is silent, instead of guessing. This matters more than it sounds. The 1850 census lists every household member by name but does not state relationships. A helpful AI may supply them anyway. The 1880 census is the first to record each person's relationship to the head of household, so on an 1880 page, that column is genuinely stated.
New skill, same formula: put the table columns in the Format, and put these three qualifiers in the Guardrail.
Step One: Sift the Clues
Each track below comes with a short practice text so you can test the prompt against a known answer. The practice texts are invented for this lesson, and every name and detail in them is fictional. Each prompt is labeled by formula part, just like Lesson 2.
BEGINNER TRACK: turn a household census into a clean list
Goal: turn a dense, unformatted household entry into a clean table of people, ages, birthplaces, and stated relationships.
Why it matters: when everything is run together, it's easy to overlook a servant, a boarder, or a middle initial, and those are the people who change the story.
Prompt:
ROLE
Act as a patient, meticulous genealogy assistant.
CONTEXT + TASK
Example: Below is a typed transcription of the 1880 census household of the Harlow family in Worcester, Massachusetts. Extract every person mentioned.
FORMAT
Present a table with these columns: Name as written | Age | Birthplace | Relationship to head of household | Exact words from the text.
GUARDRAIL
Use only what the text says. If something is not stated, write 'not stated.' Do not assume or invent relationships, and keep spellings exactly as written.
ADDITIONAL TEXT
Practice census (invented): Household 112, Worcester, Massachusetts, 1880. Harlow, Henry, head, 46, farmer, born Massachusetts. Harlow, Sarah A., wife, 43, keeping house, born Massachusetts. Harlow, George H., son, 19, works on farm, born Massachusetts. Harlow, Ellen M., daughter, 15, at school, born Massachusetts. Doyle, Bridget, servant, 22, born Ireland. Pratt, Joseph, boarder, 67, retired.
Optional: add one sentence from your Town Profile, such as the town and year, to give the AI context. For example: "Worcester was a city in Worcester County in 1880."
Answer key: a good reply lists six people. Henry is the head, and Sarah, George, and Ellen are stated as wife, son, and daughter. Bridget Doyle is a servant and Joseph Pratt is a boarder. Joseph's birthplace should read "not stated." If the AI calls Bridget or Joseph a relative, or fills in a birthplace for Joseph, it is guessing.
INTERMEDIATE TRACK: sift a will or deed and spot the FAN network
Goal: pull every named person, their role, and any property or money tied to them out of an old legal document, and flag possible FAN connections (Friends, Associates, and Neighbors).
Why it matters: in any time or place, the people standing beside your ancestor in the records often hold the key. Witnesses, appraisers, bondsmen, baptismal sponsors, adjacent landowners, and neighbors on a census page can point to a daughter's married name, a hidden relationship, or the place a family came from. Wills are our example here, but the same method works on parish registers, deeds, city directories, and passenger lists.
Prompt:
ROLE
Act as a professional genealogist and archival researcher specializing in 17th- and 18th-century New England wills and land records.
CONTEXT + TASK
Example: Below are my checked Town Profile for Lancaster, Massachusetts, and a transcription of a will from 1710. Use the checked profile to interpret place names and jurisdictions. Extract every named person with their role (such as heir, executor, witness, or adjacent landowner), their relationship to the testator, and any property or money tied to them. Then list possible FAN connections: people who appear in more than one role or alongside family members.
FORMAT
Present a table with these columns: Name as written | Role | Relationship (stated only) | Property or money | Exact words from the text. Then add a short list titled 'FAN leads to follow.'
GUARDRAIL
Use only what the text says, and write 'not stated' where it is silent. Keep original spellings and note any variants. Do not assume a witness or neighbor is a relative. Flag any word you are unsure you read correctly.
ADDITIONAL TEXT
Practice Town Profile: [paste the checked Lancaster Town Profile you built in Lesson 2]
Practice will (invented, set in our running example town): In the Name of God Amen. I Josiah Calder of Lancaster in the Province of the Massachusets Bay in New England, husbandman, being weak of body, doe make this my last Will. Imprimis, I give unto my loving wife Hannah the use of one third part of my housing and lands during her widowhood. Item, I give unto my eldest son Ebenezer my homestead lott on the north side of the river, bounded east by the land of Thomas Marsh and west by the land of Jonas Hale, he paying unto his sister Marcy ten pounds within one year after my decease. Item, I give unto my daughter Mercy, wife of Samuel Thayer, the sum of ten pounds. And I appoint my son Ebenezer sole Executor. Signed in presence of Thomas Marsh, Seth Pratt, Mary Hale.
Answer key: a strong reply finds nine people.
- Josiah Calder is the testator. Hannah is his wife, with the use of one third of the housing and lands during her widowhood.
- Ebenezer is the eldest son, the homestead lot heir, and the executor, and he owes his sister ten pounds.
- Mercy is a daughter and the wife of Samuel Thayer, with ten pounds. The text also says "his sister Marcy." A strong reply keeps both spellings, says they probably refer to the same person, and labels that an inference rather than a stated fact.
- Thomas Marsh is the neighbor on the east and also a witness, which is a FAN lead. Jonas Hale is the neighbor on the west. Seth Pratt is a witness.
- Mary Hale is a witness who shares a surname with the neighbor Jonas Hale. That is a lead worth following, but not proof of kinship, so the relationship should read "not stated."
INTERMEDIATE TRACK EXTRA: about land descriptions
Many deeds and wills describe land by landmarks, directions, and neighbors rather than a grid. These "metes and bounds" descriptions turn up in many places that inherited English land law, and the adjoining owners named in them are prime FAN leads.
AI can restate such a description as a numbered list, which makes it far easier to follow. It is poor at geometry and measurements, though, and as in Lesson 2, you should not ask it to draw anything. Ask it to restate, and nothing more. Try it on the boundary lines in Josiah Calder's will.
Prompt:
ROLE
Act as a professional genealogist and land records specialist.
CONTEXT + TASK
Below is a transcribed land description from [Town/City, State or Country], [Year]. Restate it as a numbered list of boundary calls, in order. Use my checked Town Profile, pasted below, to interpret place names.
FORMAT
For each call, give: Direction | Distance, exactly as written | Landmark or neighboring owner | Exact words from the text. Then list every neighbor named.
GUARDRAIL
Do not convert units, calculate acreage, or describe the shape. Use only what the text says, and write 'not stated' where it is silent. Flag any word you are unsure you read correctly.
ADDITIONAL TEXT
Land description: [insert transcribed land description from a will]
Town Profile: [paste your checked Town Profile]
Then compare the result with a real map. Use your Town Profile and your "Map sources" list from Lesson 2 to see whether the landmarks and neighbors make sense for that place and period.
After You Press Enter: The Reality Check
Just because the AI turned a messy record into a shiny table doesn't mean it read every name correctly. Before the table goes into your notes, run the same checks from Lessons 1 and 2, now tuned for extraction.
BEGINNER TRACK: check your table
Checks:
- Search for the quote. Pick two rows and use your device's search (Ctrl+F or Cmd+F on a computer, or Find in the page or document on a phone or tablet) to find the "exact words" in your original text. If they aren't there, the row can't be trusted.
- Spot-check the spellings. Compare a few tricky names in the table with the original. AI sometimes "normalizes" an old spelling, for example turning "Marcy" into "Mercy," when you need the spelling preserved.
- Keep the original beside the table. Save the source text and the table together, and tag the table "working copy, unverified," so you can always re-check the raw evidence. A table stays a working copy until you've checked it against the original.
INTERMEDIATE TRACK: check your table
Checks:
- Search for the quote. Pick two rows and use your device's search (Ctrl+F or Cmd+F on a computer, or Find in the page or document on a phone or tablet) to find the "exact words" in your original text. If they aren't there, the row can't be trusted.
- Spot-check the spellings. Compare a few tricky names in the table with the original. AI sometimes "normalizes" an old spelling, for example turning "Marcy" into "Mercy," when you need the spelling preserved.
- Keep the original beside the table. Save the source text and the table together, and tag the table "working copy, unverified," so you can always re-check the raw evidence. A table stays a working copy until you've checked it against the original.
- Check every relationship. Make sure the AI didn't promote an executor, a witness, or a neighbor into a biological child. If a relationship is not stated, it stays "not stated."
- Treat FAN leads as leads. A shared surname or a repeat appearance is a reason to look further, not a conclusion. Write down each lead, and pursue it in other records.
Save to Your Notes: Lesson 3
Beginner: the practice census household text and your clean table, side by side, tagged "working copy, unverified."
Intermediate: the practice will and your clean table, side by side, plus your numbered restatement of the land description if you tried the extra, all tagged "working copy, unverified."
Your Takeaway Prompts
Copy, paste, and fill in the bracketed blanks with your own details. Try the practice texts first, check the results against the answer keys, and then try the prompts on your own records whenever you like. Keep that work in its own document. Lesson 4 builds only on the practice texts.
BEGINNER TRACK: sifting a census household
Prompt:
Act as a patient, meticulous genealogy assistant. Below is a typed transcription of a census household from [Town/City, State or Country] in [Year]. Extract every person mentioned. Present a table with these columns: Name as written | Age | Birthplace | Relationship to head of household | Exact words from the text. Use only what the text says. If something is not stated, write 'not stated.' Do not assume or invent relationships, and keep spellings exactly as written. Here is the text: [Paste text here]
INTERMEDIATE TRACK: sifting a will or deed
Prompt:
Act as a professional genealogist and archival researcher specializing in [Region and Period, e.g., 17th- and 18th-century New England] wills, probate, and land records. Below are my checked Town Profile for [Town/City, State or Country] and a transcription of a [will, deed, or inventory] from [Year]. Use the profile to interpret place names and jurisdictions. Extract every named person with their role (such as heir, executor, witness, appraiser, bondsman, or adjacent landowner), their relationship to the main person, and any property or money tied to them. Then list possible FAN connections: people who appear in more than one role or alongside family members. Present a table with these columns: Name as written | Role | Relationship (stated only) | Property or money | Exact words from the text. Then add a short list titled 'FAN leads to follow.' Use only what the text says, and write 'not stated' where it is silent. Keep original spellings and note any variants. Do not assume a witness or neighbor is a relative. Flag any word you are unsure you read correctly. Town Profile: [Paste your checked Town Profile]. Text: [Paste text here].
Coming up in Lesson 4:
You've mapped your place, and you've turned messy records into clean, checkable tables. But once that evidence is sitting in your notes, how do you cite each record so you, or a cousin, can find it again?
In Lesson 4, we'll use AI to draft working citations from source details, flag what's missing, and test a published family history's claim against the original record. You'll also learn how to catch the mistakes AI makes along the way.
Bring to Lesson 4:
- Beginner: the practice census household text and your clean table, tagged "working copy, unverified."
- Intermediate: the practice will, your clean table, and your numbered restatement of the land description if you tried the extra, all tagged "working copy, unverified."
What is the messiest, most difficult record you've ever had to transcribe by hand? Drop a comment below!
Found this helpful? Share it with a fellow researcher: