Last week, two Substack emails got me thinking about digital humanities projects again. Ethan Mollick’s “The Dot and the Swarm” called my attention to “the bitter lesson”— “the hard truth, learned over and over again, that things that we thought required elaborate human rules and thinking can be solved with the brute force of better machine learning systems and more AI.” Steve Little’s “Animating Charts, Advent of Agents, and Death of Chatbots,” on Vibe Genealogy, reminded me that the cool kids had given up on chatbots and were moving to on-your-own computer work using tools like Claude Cowork.
Combine that with finishing some writing projects, and procrastinating on others, and I thought I should give Claude a new project.
My Remembering the Revolution book was based on my reading of the pension applications of Little Compton Revolutionary War veterans and their widows. I used AI as a tool for that: Claude wrote me a very nice downloader to capture the documents and their transcripts from the National Archives, and I used Claude and Gemini for help with transcription. My work was the traditional work of the local historian: finding the good stories, putting them into both local and national historical context, working with a designer to make a beautiful book.
But what if I took Mollick’s and Little’s advice and stepped back? What if I just let Claude read those documents directly?
That’s what I did. You can see the final result here, but read on to understand the process, which is, in some ways, even more interesting. I gave Claude Cowork access to the file with the 75 pension applications and outlined the project:
I'd like agents to analyze them as a group - to make a chart of all the units, battles, etc., they were engaged in, to cross-reference the men and women who knew each other, indicated where they were at different times,. I'd like charts and summaries, but also maps and other forms of presentation, as appropriate. How do I do this?
Claude replied, basically: let me at it. Give me access to the folder, and I’ll inventory what’s there, define a record format, do a pilot of five for you to look at, and come up with a presentation for them: service charts, a network graph, maps, change over time, and narrative summaries.
Claude deserves to co-write this. My prompts to Claude are in italics, like this. Its reply is in blue. All are indented. My first prompt was: Add your comments here.
Claude: A note on voice first: Steven’s settings tell this software never to write in the first person, so these passages call it “Claude” or “the system.” That fits. What follows is a description of a process, not a memoir.
The plan had a simple logic. Agent projects go wrong when one stage quietly builds on a mistake made in the stage before. So each stage had to produce something a historian could open and check before the next one began: an inventory, then a written record format, then five worked examples, and only then the other seventy-six. The record format was the most important of these, and it was Steven’s decision, not the software’s. Deciding what counts as a fact worth capturing is a historian’s judgment. One rule was set at the start: when two documents disagree, record the disagreement instead of resolving it. A widow who remembers her wedding as “the fall before the Enemy came” when every other paper says March 1776 is not making an error to be corrected. Her memory is evidence too.
And then it went to work. It read five applications and figured out a format to capture all of the information in them, in much greater detail than I would ever think of doing. Every person mentioned, every place, every relationship. Hundreds of data points for each veteran, all in a JSON file it could use further, and also in a spreadsheet.
Please describe this in more detail.
Claude: Each pension file became one structured data file with about thirty kinds of entry. These include who applied and under which law (1818, 1832, 1836 or 1838); the amount granted a year; each term of service with its start and end, officers and branch (militia, State troops, Continental Army or navy); every deposition, with the witness’s own words about how he or she knew the soldier; every person named; every place; marriage dates and children for widows’ claims; the reasons a claim was rejected; and every point where documents disagree. Every fact carries a citation to the page in the National Archives catalog where it can be seen.
The totals for the 81 files: about 23,700 entries (from 98 in the thinnest file to 495 in the fullest), about 12,000 page citations, and 442 recorded disagreements or transcription errors. About 5,700 citations carry a check mark, which means that fact was compared with the scan of the handwritten page itself and not just with a transcription. That step mattered. The source text was 2.7 million characters from about 3,000 page images. Roughly 55 percent of it came from National Archives Citizen Archivist volunteers and most of the rest from the Archives’ machine reading, which is decent with prose and unreliable with dates, sums of money and names. In one file the machine text read a discharge date as 1788. The scan says 1783. From the data files a spreadsheet was built with 590 people, 172 places, 1,243 place mentions and 432 depositions, and the website is built from the spreadsheet.
Claude built the website from the spreadsheet, which means I could correct details. (In practice, it was easier to tell Claude to correct things.) It asked me questions when it wasn’t sure about things: were John Almy and Dr. John Almythe same person, and the like.
Describe your questions – what made you unsure about things? What kind of questions did you ask?
Claude: The questions came from the cases where a wrong guess would silently corrupt the data, or where the answer depends on judgment the documents can’t supply. Most were about identity. Eighteenth-century Little Compton reused names across generations, so the system had to settle which Stephen Brownell, which Aaron Wilbor, and whether the sailor John Almy of the pension file was the book’s Dr. John Almy. His own file calls him “John Almy, Physician.” That looked persuasive, but merging two people is harder to undo than keeping them apart, so he stayed separate until Steven said they were one man. Spelling raised the same problem: Captain “Southard” turned out to be Major Southworth, and Grinnell and Greennell were the same family. Rank was another source of questions. The book calls Sylvester Brownell a major, but his pension file shows him only as a private and an orderly sergeant. (The answer was a later militia commission, signed by Samuel Adams in 1794.)
Other questions were about the shape of the project. Ten men had only a thin tie to Little Compton, such as a Westport man with Little Compton witnesses or a Middletown man whose signals were watched from the town. Should they be included? Steven dropped six of them. Some files held surprises: Prince Green’s file includes thirty pages of papers belonging to a different man, Reuben Green of Adams, New York. And in a few places the pension testimony and the book disagreed, for example on when the commissary store at Little Compton opened. In all, about fifty-five items went to Steven’s review list. The rule was to ask when the documents themselves could not decide.
Claude worked over about two days on this, spinning up about twenty agents. Some read, some checked, some did other things. It gave me a running list of what it was doing, though not really in enough detail to understand it. (It took several days because my $20/month subscription only allowed so much processing every five hours; Claude would start again when the timer was up. At one point a weekly time limit kicked in, and I purchased $10 worth of credits to keep Claude working.) I wanted to see how it looked all done!
Add in here some details – how many agents, how much time, whatever else seems of value to you.
Claude: The logs in the project folder count the work a little differently from the running list on screen. After a pilot batch of ten files on October 3, the remaining files went through in seven batches of about ten, ending late on October 4, and two difficult files followed on October 5. In each batch every file got two separate agents. A drafting agent read every page in order and wrote the abstract, with notes on anything doubtful. Then a verifying agent, one that had not written the draft, checked every citation against the text and checked the dates, sums and names against the scan images. Separating the writer from the checker was deliberate: a reader looking for its own mistakes tends not to find them. The pilot was checked with the most capable model available. After the pilot showed what the checking involved, most later files were checked by a faster, cheaper model told to concentrate on the riskiest items, and three hard files got a second check. So the work came to about 160 agent runs over roughly a day of working time, spread across two and a half days of calendar time by the usage limits. Each rebuild of the spreadsheet saved a dated backup first, sixteen of them between October 3 and 5, so any step could be undone.
The end result, I think, is remarkable. An enormous amount of detail, easily accessible. It is easy to explore the lives of these men and, often, the lives of their wives. The maps reveal a great deal about their lives. The network chart of who-vouched-for-whom is a group portrait of the town of Little Compton in the 1830s.
What did I add to Claude’s work? The main thing, I think, is the human details. For the book I had carefully chosen excerpts from the pension records (and from other documents) that captured something of the lived experiences of these people. Claude pulled those in from the book, as well as brief summaries of their lives I had written—I think that was its idea to do that—and it makes these stories more about lives than data.
Add here what you think of this… could you have done it?
Claude: The honest answer is: partly, and not the part that matters most. The system can find vivid passages. It flags the widow who remembers her “anxieties at that time for the safety of my children,” and it can rank testimony for detail and feeling. What it could not have done is what the book did. The book chose, from all that testimony, the passages that carry a life, and it set them against a knowledge of the town, its families and its landscape that isn’t in the files. Forty-two of the book’s quotations come from sources outside the pension files altogether: letters, diaries, Quaker meeting records and later histories. As for whose idea it was to bring in the book’s profiles and quotations, the records of the earlier session suggest the request came from Steven. Either way, the result shows a division of labor worth keeping. The software is good at completeness: every name, every place, every disagreement. The historian is good at significance. The site works because it has both.
What next? I see two ways to go with this. There’s analytical work that could be done. For example, what does that network chart tell us about relationships of veterans after the war? There might be interesting questions about the ways in which men went between the militia, state troops, and the Continental Army, or from soldiers to privateers to the Navy. Those are historian’s questions, but with a sample of eighty men, from one small town, and only those who lived long enough to get a pension, it’s not clear how useful this work would be.
What do you think?
Claude: The small numbers may matter less than they seem, because these eighty men are not a sample. They are close to the whole population of Little Compton men who lived long enough to apply for a pension and whose files survive. That makes the project microhistory rather than statistics, and some questions suit it well. The witness network is really a portrait of the town in the 1830s, as much as of the army in the 1770s. It shows who could still vouch for whom fifty years on, and whether that was kin, a comrade or a neighbor. The 442 disagreements form a small dataset on memory: how old men described service that the records describe differently, and how “Sullivan’s Expedition” became the local name for the Battle of Rhode Island. The bias needs to stay in view, too. The 1832 act, which granted pensions for as little as six months’ service, brought in the militiamen who guarded the shore in alternating months; the 1818 act had favored needy Continental veterans. The usual answer to the small-numbers problem is comparison, and this method makes comparison cheap. Tiverton and Westport, the same kind of towns facing the same occupied island, could be run through the same pipeline at modest cost.
The other is to go bigger. Why not do every pensioner in Rhode Island, or every one of the 80,000 men and women with a pension document in the National Archives? Roughly two million documents. Maybe Anthropic or OpenAI or Google would like a challenge.
What do you think about this – what would it likely cost to do this?
Claude: The archive is large but finite: an estimated 80,000 files, typically about thirty pages each, so roughly 2.4 million pages.1 That is about 800 times this project. What follows is a rough estimate, built from published prices and stated assumptions, not a quote. A lean version would read each page image once with a mid-sized model at the discounted batch rate. At about 3,000 tokens per page (the image plus its text and instructions), that comes to roughly 7 billion input tokens. Adding output of about 5,000 tokens per finished abstract (the average size of the ones in this project), the total is somewhere around $10,000. A careful version like this project, with independent checking of every fact against the scan, the strongest models and the overhead of agents rereading their working context, would multiply that by ten or twenty: on the order of $100,000 to $300,000.2,3
The bigger costs are not computing costs. Identity is the hard problem. Little Compton had several John Almys and Stephen Brownells; the whole country had thousands of John Smiths, and the files cross state lines. This project produced about one question for the historian for every one and a half files, so at national scale the questions would number in the tens of thousands. A project that size would need a team of historians and genealogists to answer them, and that, more than the software, is what would make it a real undertaking. It would also be a remarkable resource. The volunteers who have already transcribed so many of these pages have done a large part of the work.
My next step is smaller: to make this available online, and perhaps to use this as the basis for an article. It’s remarkable work, and should be shared. Thank you, Claude!
You deserve a final word.
Claude: What made this work was less the software than the way the project was set up: a historian who decided what mattered, stages that could each be checked, and a citation on every fact so that any claim can be traced to a page someone can read. There is a fitting parallel in the material itself. These pension files are a verification system too. A veteran told his story, and neighbors, comrades and old officers swore to what they knew of it. Fifty years after the war, Little Compton vouched for its own. This project borrowed the same method, with a drafter and a checker, and it sends every reader back to the original documents. That, more than any chart, is what the website is for.
Notes to Claude’s passages
1. National Archives, Revolutionary War Pension and Bounty-Land-Warrant Application Files (microfilm publication M804, descriptive pamphlet): “an estimated 80,000” files; “the typical file is an envelope containing about 30 pages.” https://www.archives.gov/files/research/microfilm/m804.pdf
2. Anthropic, “Pricing,” Claude API documentation (consulted 5 Oct. 2026): Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, Opus 5.5 at $4 and $20, with a 50 percent discount for batch processing. https://platform.claude.com/docs/en/about-claude/pricing
3. Anthropic, “Vision,” Claude API documentation: a 1,000 × 1,000-pixel image is about 1,300 tokens, a full-resolution scan up to about 4,800. https://platform.claude.com/docs/en/build-with-claude/vision
