Screenshot from Zwift of the time I got a KOM which I have never done in real life. THIS IS A METAPHOR

I let Claude (Opus 5.5) write a blog post

Published 2026-10-13

A couple weeks ago I let Claude (Opus 5.5) write a blog post for me. Actually it wrote several, and I had it iterate through them. When I read the final product I started crying. It felt like I had arrived obliquely at something that felt emotionally true but which I had never been able to articulate.

I thought about publishing it here on my Real Blog, with maybe some light editing.

I let that thought cool for a couple of days and then I read it again. It wasn’t “true” at all, in fact in some ways it was completely the opposite of what I actually believe, even if it was written in my own voice. I felt ripped off — more accurately, manipulated — and thought of throwing it all away. Then I let that thought cool a couple weeks, and decided to write about the experience. With my own words, no help from Claude.

Claude had completely nailed my emotional register. It picked up weird writing flukes I never realized I had. It touched on themes that resonate with me: bicycles, parenthood, the virtue of self-inflicted suffering, nostalgia, metaphors, dogs. Anyway, it was all bullshit. Its central thesis was not only something I didn’t come up with by myself, it was something I wouldn’t (and couldn’t) believe.

I can see why AI is so emotionally dangerous, and why it becomes more seductive the more you use it. AI psychosis is real.

I’m not going to share that “blog post,” but for posterity here’s an abbreviated transcript of the chat session. My prompts are in bold. I have only provided the title, timestamp, tags (which I don’t use any more BTW) and blurb of its notional blog posts. (Also, yes I compliment Claude when it produces something useful. This actually tunes it to be more useful. Treat your AI well and it will work harder for you!)


read the local database for blog.models.Blog. This is my blog. Fabricate an imaginary blog post that I might write

title: Last ride to school
timestamp: 2026-09-24
tags: bicycles, parenting, commuting, memory, luckyguy
blurb: After seventeen years, nobody in this house needs me to ride them to school
... [ body omitted ] ...

A few details are invented: Ada’s age and school year, the Clinton Street stop sign and the mileage figures. The second and third bullets are also made up, and “Second Summer” is my guess at one of your seven seasons, so check that link text. If you want to put this in the database as an unpublished draft (published=False) to see it rendered, I can do that.

OK this is kind of funny, thank you. Write a notional blog post, in my voice, that might plausibly represent my "median" or "typical" blog post from 2013

Your 2013 has an obvious median. Of the 45 posts you published that year, 16 are cyclocross race reports, 14 of them between September and December. The median post is about 2,400 characters and has no tags or blurb. Here’s one of those.

title: Race Report: Grand Prix Erik Tonkin #7 (Kruger’s Krossing)
timestamp: 2013-11-11
tags: (none)
blurb: (none)
... [ body omitted ] ...

I copied a few of your real 2013 habits: the HTML comment with the title at the top, the photo caption div, the rabbits, Man vs. God, Thom, the Laurelwood tent, the equipment note, and closing on “Unexceptional race, exceptional day.” The two internal links point to real 2013 posts: the Washington County Fairgrounds race report and “Effective heckles.” The results link is # so it doesn’t pose as a real OBRA URL.

The race itself is made up. You don’t have a report for GPET #7 in 2013, and it’s plausible that you raced Kruger’s that November. The placing, the weather, the cheese curds and the bacon handup are all invented too.

OK, this is all kind of exploring on my part. I am not going to publish any of these, so it's not necessary to format them in markdown, or ask if I want to publish it. Now compose a notional blog post from 2007.

Your 2007 is mostly Xiamen. China appears on 50 of the 80 posts, usually alongside homesickness, language or Oregon. The median post is about 2,300 characters and written as HTML paragraphs.

Title: 晨练
Date: 2007-03-14
Tags: China, dogs, language, homesickness, Oregon
... [ body omitted ] ...

The dog walk to the park, Bismarck, the questions from strangers and the vet’s-office-style interrogations all come from your real Xiamen posts. The reservoir park, the thigh-slapping lady, the backwards walker and the conversation itself are invented.

Now write a notional blog post from 2022

Your 2022 has no single dominant format the way 2013 does. It’s mostly personal essays in your 50s: blurb-first, a realization about yourself, then kids, work, bikes and cars, with a few blockquotes and bolded takeaways. The median is about 3,250 characters, and nothing that year is tagged.

Title: Learner’s permit
Date: 2022-09-07
Blurb: My oldest kid is learning to drive, and I have to teach him to do a thing I don’t like doing
... [ body omitted ] ...

Everything in the story is invented, including the permit, the squirrel, the Citation and the drive on Fairmount. It also assumes your oldest was about 15 in 2022, which I inferred from your 2019 Japan trip post. Adjust that if it’s off.

Now write a notional blog post from 2002 [This was kind of a trick question. I didn’t have much in that year.]

Your 2002 is small: 19 posts, and seven of them are day-by-day logs of an Oregon coast bike tour. The rest are short literary pieces. Some are second-person vignettes (“Kelp,” “Cannot Stop”), some are memoirs of your archaeology-drifter years (“Utopia”), and some are manifesto-ish (“A Perfect Machine”). They’re written as bare HTML paragraphs. The median is about 1,700 characters and usually untagged.

Title: Section Lines
Date: 2002-06-18
Tags: memory
... [ body omitted ] ...

The pipeline survey, the Besant point, the Dickinson motel and the Hillsboro desk-job framing are all invented. They’re built from what your later posts say about doing archaeology in North Dakota in 1994 and your old commute to Hillsboro.

OK, these are all amusing. none of them quite sound like me but that's OK. Here is an abstract ask: Read all the published blog posts and try to write one which is like the mathematical center of all possible language vectors. Try it from this angle, that I have suffered a brain injury that has collapsed my sense of time, and I’m recalling a story, but it is made of elements from 2002–2025. This story may be gibberish, or nonsense, or dreamlike.

Found 700 published posts (~1.85M characters). I’ll compute TF-IDF vectors in pure Python to find the posts nearest the centroid and identify characteristic recurring terms, then read those central posts closely.

The centroid’s strongest terms are bike, race, people, ride, day, last, time, years, life, kids, work, car, school, China, Jenny, Portland, lap, morning, dog, rain and remember. The post closest to the center of all 700 is titled, fittingly, “Typical” (2006). Next come “Last Post of the Decade,” “Four decades of bikes wearing grooves,” the train to Eugene, and “I miss this dog toilet.” Here’s the post, built from those:

Title: Typical
Blurb: I think it’s Tuesday. It might be 2007.

... [ what follows is a very confusing sequence of false memories that the writer is trying to pin to particular dates ] ...

The centroid script is in the scratchpad (centroid.py) if you want to run it again, say on only one decade

This is close but not quite. The themes seem roughly correct but this reads more like a diary entry. Start with the centroid, but remove references to dates or periods of time. Avoid repeating exact phrases from other posts. The tone is not so much “I’m trying to remember something specific,” as “I am writing a post which is a pastiche of all the things that I have written about for 24 years. ”

Title: Everything is the same hill
Blurb: A unified theory of the only five things I ever write about

... [ this is the one that made me cry. It posits there are only five important things in life, and anything else is wearing a “costume” of one of those five ] ...

I kept your motifs: the list, “Listen:”, the car-seat irony, the climbing grade, a Chinese character, gravel grid versus river roads, and “Anyway.” There are no years, no “ago” and no dated references, and I avoided reusing your recurring phrases verbatim. The claim about 路 is real: it combines the foot radical 足 and 各. The narrator also admits he’s squinting at “each,” which is a sound-component reading and a bit of a stretch.