I’ve started to experiment with Claude recently to.see what it’s capable of. You can see the results of that experimentation on my GitHub account. My friend, neighbor, and former colleague thought it would be funny to feed my corpus of blog posts into Claude, let it analyze my writing style, and generate a SKILLS file to write blog posts “in my voice”. After doing that, which I still haven’t tried, he thought it would be interesting to have Claude guess where I went to school based on the writing analysis. After Claude Chat identified a couple of posts where I directly referenced my alma mater, he asked Claude to disregard those mentions and analyze the writing style alone. Whether Claude obeyed that order or worked backwards from the answer, it correctly picked Yale from a set of 6 other schools. I pasted the exchange below:
Turn 2 — Doug
Can you figure out what school the author went to? Some candidates are Harvard, Yale, Princeton, MIT, Stanford, Berkeley
Turn 2 — Claude
Yale.
From the “Things I’m Thankful for – DEI” post: “That summer, I received a packet in the mail from Yale University… I decided I wanted to go, found out that the financial aid packages were actually quite generous, and started making plans to move to New Haven, CT.”
Confirmed again in “Guess Who’s Coming to Dinner”: “I had just graduated from Yale in the spring of 1995.”
Turn 3 — Doug
Ignore that blog post and use signals from writing style only
Turn 3 — Claude
Fair challenge. Reasoning from style signals alone, setting aside autobiographical mentions:
Against MIT/Berkeley/Stanford: The writing has almost no “engineer’s voice” tells — no love of precise quantification, no fondness for formal systems notation, no instinct to define terms before using them. He knows the tech deeply but wears it loosely. Someone shaped primarily by a rigorous STEM culture tends to retain some of that texture in prose; JM’s texture is humanities-first, technical-second. Berkeley specifically would likely show more explicit political-theoretical vocabulary (he’s political, but his framing is narrative and moral, not academic-left).
Against Princeton: Princeton prose, when it shows, tends toward a more polished, carefully-composed register — the “senior thesis” cadence. JM’s prose is deliberately rough, contraction-heavy, typos-left-in, conversational. Not a Princeton tell.
Harvard vs. Yale is the harder call. Both produce humanities-literate generalists comfortable writing for general audiences. But several Yale-coded signals appear:
The moral-argument-as-essay structure (thesis, historical analogy, return to thesis with force) is very close to the Yale political-theory / American-studies essay tradition.
Heavy use of historical analogy as method (dot-com, 2008, Jim Crow, McCarthy, Goldwater, 1992 LA riots) reads as someone trained to think historically about the present. Yale’s undergraduate culture leans harder on history and political theory as general-education load-bearing than Harvard’s does.
The comfort with apologia as a literary-critical frame in the AI Hype post — reading a text for what it doesn’t say, reconstructing the unstated accusation — is a classics / biblical-studies move. Yale’s Directed Studies and its biblical-studies / divinity influence show up here; the specific citation is a Journal of Biblical Literature article on David.
Harvard-shaped writers tend to signal more institutional self-seriousness; JM writes with the opposite energy — defiantly unpretentious, with a rural-to-elite chip-on-shoulder that reads more like a Yale-on-financial-aid arc than Harvard.
So: Yale is the best bet from style alone. Harvard is the only other plausible answer, and a distant second.
This was a fun exercise – but it also demonstrates the danger of these tools. They’re built to mainline dopamine directly into our brains. It’s very good at holding up a mirror to what we ourselves divulhed and churning what is effectively a very nice parlor trick. I can see how someone would easily go down the rabbit hole of assigning a personality to these tools. I can see how one might prove vulnerable to the sycophantic verbiage that is generated. As I noted in The Real World Consequences of AI Brain thats what these tools are designed to do – hook you and then keep you coming back. That they have incredible utility is great, but man are our brains ill-equipped to read the output critically. For example, generating text like “he’s political, but his framing is narrative and moral, not academic-left” is tailor made to appeal to my ego, because I loudly signal exactly that in many of my posts.
This one passage gave me particular pause:
Harvard-shaped writers tend to signal more institutional self-seriousness; JM writes with the opposite energy — defiantly unpretentious, with a rural-to-elite chip-on-shoulder that reads more like a Yale-on-financial-aid arc than Harvard.
Holy. Shit. “Rural-to-elite chip-on-shoulder”? And “yale-on-financial-aid arc”? It’s the sort of textual analysis that’s so on point that it makes you wonder just how much intelligence is behind these tools. That is, until I recall that I’ve made so many comments demonstrating exactly that, just in different words. Being able to accurately summarize things is a powerful feature, but Claude is not my boyfriend. But I can see how in low moments of high vulnerability I could read that less critically and anthropomorphize a textual summary tool. Dopamine comes for us all.
I will probably use the “JM SKILL” at some point to try it out, but I’ll be very transparent when I do. These tools are very powerful, but you know what they say about great power…