How this site was actually made

What I bring to this and what I don’t

My background is academic legal. I teach at Vanderbilt Law School, and before that I was a law librarian. I am not a digital humanist, a corpus linguist, or a computational researcher by training. I claim no expertise in those disciplines. What I brought to this project is the editorial habit of a lawyer-turned-librarian: care about sources, attention to attribution, and a long working familiarity with how research literatures are constructed. The librarian background in particular helped: a good portion of the work was knowing what to ask for, how to verify it, and where to look when something didn’t hold up.

That said, this is a living project. It may contain mistakes. Some will be mine, in framing or interpretation; some will be the AI’s, in the underlying analysis or in the prose. The data, the analysis scripts, and the site code are all open under permissive terms. I would much rather have someone notice an error and tell me than present the work as settled. The right way to read this site is as a starting point for further exploration and contribution, not as an endpoint.

What I did

I posed the original question: did the Founders sound like Shakespeare, and could you tell from the actual corpus rather than from received opinion? I chose the corpora (the Founders Online archive on one side, the Project Gutenberg Shakespeare on the other), and I chose the methodological framework, Anatol Stefanowitsch’s open-access Corpus Linguistics: A Guide to the Methodology (Language Science Press, 2020), as the spine the project would hang on.

I made the editorial calls. Which findings rose to a case study, which ones stayed in the supplementary paper, which framings worked for a generalist audience and which were academic-only. I corrected mistakes when I saw them. The project went through several revisions, often substantial ones, because something Claude generated turned out to be wrong on a date, on a recipient, on the underlying statistical claim, or on the rhetorical register. I rejected pages of prose because they felt wrong. At one point I asked Claude to soften every page for a general reader and move the statistical apparatus behind expandable panels; the biographical-scenes-first orientation you see now came out of that pass.

I read every page carefully. The site has many specific historical claims in it: that Adams wrote a particular letter to a particular recipient on a particular date, that a particular Shakespearean passage comes from a particular act and scene, that a particular Founder used a particular word at a particular per-million rate. Each of those claims either traced to a row in a CSV file or to a verifiable archival reference. The ones that didn’t hold up didn’t stay on the site. I am sure some still slipped through. If you spot one, please open an issue on GitHub.

What Claude Code did

Most of the production work. The Python ingest pipeline that pulled the Founders Online and Project Gutenberg corpora into the SQLite database. The analysis scripts implementing the eight case studies, the three influence-ranking measures, and the catalogue and metaphor pipelines. The normalization layer that harmonized Shakespeare’s First Folio spelling with the Founders’ eighteenth-century usage. The export step that produced the JSON files the site reads.

And the site itself. Every page template. The Next.js scaffolding. The Tailwind design tokens for the parchment palette and the EB Garamond / IM Fell typography. Every interactive explorer: the Honour Test, the Catalogue, the Quotation Timeline, the Ranking matrix, the Metaphor Fingerprints, the Archaic Threshold, the Play Atlas, and the others. The SVG charts. The GitHub Pages deployment. The Founders Online and Folger Shakespeare backlinks. The image-credits attribution table.

And most of the prose. The essays in The Commentary, the case studies in their full narrative form, the methodology explanations in plain English, the closing arguments. I prompted, redirected, corrected, sometimes rewrote, sometimes accepted with light edits. But the bulk of the words on this site were drafted by Claude.

What that means for how to read the work

The findings exist independently of whoever wrote them up. The Python pipeline counts what it counts. Macbeth occurs in Adams’s 1758 diary the number of times it occurs there. Hamilton’s 2.21 million words contain zero named Shakespeare references at the catalogue’s confidence threshold whether a human or a machine reports the count. The differential collocates of honour in the two corpora are what they are. The eleven-method convergence matrix says what it says.

What the AI collaboration does change is the standard of attribution. This is not a piece of solo human scholarship that happened to use spell-check. It is a piece of work whose Python implementation, prose drafting, and visual design were substantially the product of an AI coding agent operating under a human editorial hand. I’d rather acknowledge that up front than have a reader infer it later.

One reason this matters: a reader looking at a clean page of prose with statistics in it has reasonable grounds to assume the author has personally verified every number and every claim. In a conventional scholarly setting, that assumption is roughly correct. In this setting, the most I can honestly claim is that I have verified every claim I noticed. Where my eye failed, an error may remain. That’s another reason the project is presented as a starting point and the underlying data is open.

What the workflow looked like, roughly

The project moved through several rough phases. I don’t want to overclaim a tidy linear process. In practice every phase overlapped with the others and several got revisited multiple times.

1. Corpus ingest and normalization. I described the corpora I wanted, what they were good for, and what the licensing constraints were. Claude wrote the ingest scripts, ran the downloads, built the SQLite database. I checked that the document counts and word counts matched what the source archives advertised.
2. Methodological framework. I named Stefanowitsch and Gries & Paquot as the spine. Claude proposed an eight-case-study structure that matched the textbook’s typology. We iterated on which case studies to keep and which to fold together.
3. Analysis runs. Claude wrote the per-case-study Python. I read the output tables, flagged surprising or implausible numbers, and we re-ran or re-specified analyses until I trusted the results. A late methodology-review pass surfaced an honest normalization bug in one case study that, when fixed, slightly changed the per-Founder distances but not the headline ranking.
4. Paper drafting. Claude drafted the scholarly paper in Stefanowitsch’s mode. I read it as I would read a graduate student’s first complete draft, and we revised iteratively.
5. Site architecture. I sketched the site shape I wanted (commentary essays, case studies, interactive explorers), modelled on Lincoln Mullen’s America’s Public Bible. Claude built the Next.js scaffolding, the design system, the page templates, and every interactive component. We iterated on colour, typography, and spacing.
6. Public-audience rewrite. The first version of the site was too academic. I asked Claude to soften every page for a general reader while keeping the methodological detail accessible behind expandable panels. The case studies got rewritten; several explorers got renamed; the biographical-scenes-first orientation emerged in this phase.
7. Verification (ongoing). I’ve gone through the catalogue and the case studies looking for date errors, misattributions, and misquotations. I’ve found several and corrected them. I assume more remain.

What I think the experiment shows

I’m not in a position to draw strong general conclusions about AI-assisted scholarship from a single project, and I won’t pretend otherwise. But a couple of things struck me as worth flagging for anyone considering similar work.

The capability is more substantial than I expected going in. With a clear research question, a careful methodological frame, and editorial attention, the agent produced a complete digital-humanities project (corpus, pipeline, paper, site, interactive layer) that meets at least my standard of disciplinary care. I am sure a corpus linguist with formal training would find things to push back on, and I welcome that pushback. But the work is not slop in a wrapper.

The judgment burden does not collapse. Production cost falls dramatically, but the cost of knowing what to ask for, what to accept, and what to send back does not. Every page of this site has at least one place where Claude proposed something I rejected or had to redirect substantially. The work that remains for the human editor is the work that has always been hardest: taste, attribution, restraint, knowing when a finding isn’t actually a finding.

The category of “author” bends. I cannot honestly say that I wrote this site in the way I’ve written previous things. I can also not honestly say that Claude wrote it without me. Most of the language is Claude’s; almost none of the decisions are. The substantive findings exist independently of either of us; the framing is mine; the execution is largely Claude’s. The right word for the relationship between us is one nobody has settled yet. I’m responsible for what’s here either way.

An invitation

The site repository at github.com/willimj3/shakespeare-in-the-republic is public. The research repository contains the corpus pipeline and the analysis scripts; the papers page has the three written deliverables and direct downloads of all the underlying JSON data. Everything is under permissive terms.

If you find an error, a misattribution, a passage that’s been quoted wrong, a date that’s off, a statistical claim that doesn’t hold up, or a framing that overreaches the evidence: please open an issue or get in touch. If you want to extend the corpus, run a different analysis, or build something on top of the data, that is exactly the use the work is designed for. The point of opening the data is to let other people push the inquiry further than I could on my own.

Return to the homepage, read the substantive findings in the essays, or download the work in the papers section.