Skip to content
AI-Native PM
Human decision behind every word

How We Built This Site

We built this the way we teach you to build with AI: not in one long chat, but by directing fleets of agents, reviewing what came back, and deciding what to keep. Here is the honest record, by the numbers.

How the site was built, drawn as a loopThree boxes read left to right. We, who direct and choose, sending 400+ prompts. A fleet, which builds in parallel, running 911 agents. The site, what got made, 201 thousand words. A return arrow runs underneath from the site back to us, labelled review, correct, repeat. Caption: about one in five of our messages was a request to do it differently.THE BUILD, IN A LOOPreview · correct · repeatWEdirect and choose400+ promptsA FLEETbuild in parallel911 agentsTHE SITEwhat got made201k wordsAbout one in five of our messages was a request to do it differently.

Underneath the writing sits the machinery: 69,163 lines of code, 252 components, and a 34-rule voice spec we held every sentence to. The numbers below are grouped by who did what.

What we built

216,630
words of original writing
more than two and a half books' worth
107
teaching chapters
across 3 levels and 13 parts
206
custom diagrams
every one in the house style
14
field notes
plus 4 framework essays

How we worked

150+
days, start to here
from the first idea with Claude
400+
prompts we wrote
one conversation, start to finish
116
commits
every one our call to make
350+
hours at the desk together
ideation onward, with Claude

The fleet that did the building

~911
agents run in parallel
a fleet, not a chat
45
multi-agent workflows
fan out, then verify
~74.5M
tokens generated
the work the AI produced
~12,400
turns in the build
refreshed its memory 17 times

Staying in the loop

~1 in 5
of our messages was a correction
“simpler,” “revert,” “none of these”
~85
rewrites and redirects we asked for
the first version was rarely right
8
times we stopped the AI mid-task
the human kept the brakes
68
times we decided when to save
nothing shipped on its own

The number we keep coming back to

Of everything here, the last group matters most. About one in five of our messages during the build was a correction: “revert that, I don’t like it,” “this reads like a paragraph, tighten it,” “you didn’t hold this to our own voice rules,” “the serif is throwing me off, use our font system,” “the number says little, give me something of higher value.” The first version was rarely the right one. What made the site good was the choosing and the cutting, done over and over, which is the whole argument of Taste Is a Muscle, Not a Gift.

We wrote about the method while we were using it: How We Used Agent Fleets to Build This Site, How We Caught AI Mistakes With a Second Agent, How We Used Preview Labs to Design This Site, and How We Used Loops Instead of One-Off Prompts.