COLOPHON
An Archivist Keeps This Site Honest
I built an Archivist to look after this site. It reads my research, drafts against it, and flags what it can’t verify. I make every call.
Here’s the honest division of labor: the Archivist drafts most of the prose you’re reading on this site, working from my primary sources (the decks, transcripts, and data sitting behind each case study). It also does the jobs nobody puts on a portfolio: checking claims against source files, auditing live pages, hunting layout bugs, and keeping a log of everything it touched. I decide what’s true, what ships, and when.
As of July 2026, the machinery is Claude Code, a folder of plain markdown that doubles as an Obsidian vault, and Framer’s agent API. The pattern is borrowed from Andrej Karpathy’s LLM Wiki idea: give the agent a wiki it owns, keep the raw sources untouchable, and make it show its work.
Raw stays raw
Everything begins in a folder of primary sources: final client decks, eye-tracking exports, interview transcripts, analytics data. The Archivist reads from that layer and never writes to it. From there, everything flows one way, through a gate.
raw sources
wiki
site copy
my go-ahead
live site
Raw sources stay immutable. The wiki is the Archivist’s distilled, cross-linked notes, each page tagged with how much to trust it. Site copy is drafted in my voice with every open question attached. And the highlighted step is the gate: nothing passes it without me.
Guardrails you can see
Every one of these is a real artifact in daily use, and you’re seeing them exactly as they exist in the files.
Trust chips
Every wiki page declares how much to trust it. A claim from a draft page doesn’t ship without a second look.
complete
draft
needs-review
stub-todo
Publish blockers
Every draft ends with a checklist of what stands between it and the site. An empty list means ready to paste. They’re rarely empty.
Hero: WovenByToyota_Trailer_Final.mp4, a polished cut but not viewed in this pass. Swap it if it’s not the trailer you want.
The em-dash rule
The Archivist is banned from publishing an em-dash. That rule exists because agent-drafted prose loves them, and I’d rather confess the tell than hide it. This page is bound by the same rule.
The gate
The Archivist can stage changes to the site. Publishing requires me, every time.
Follow one number all the way down
The biggest number on this site is the revenue figure on the Shedd Aquarium study. Here is its chain of custody, stage by stage.
Raw sources
The final client deck’s GA4 analysis (March 2025): Illinois residents pay a fixed $19.99 and convert at 4.7%. Everyone else sees variable pricing averaging $41 and converts at 3.1%, despite browsing 56% more. No document in the archive computes what that gap costs. The working draft that was supposed to had literal “$X / month” placeholders where the number would go.
$95–107K/mo
computed, defended, shipped
The arc, in one line: the archive never contained the headline number, so the system computed one it could defend, and that’s the number that shipped.
Three sessions from the logbook
Real prompts, quoted from the session transcripts with my typos intact. Each ends with the part that matters: the call was mine.
The revenue number that got more honest
July 4, 2026
While drafting the Shedd study, the Archivist flagged:
The ‘six figure monthly’ headline is an extrapolation no raw document computes.
It worked the math from the GA4 data, filed the defensible range, and held the draft until I ruled.
My call: Ship the honest range. $95 to 107K is still a headline.
The bug that ate every image
July 9, 2026
The audit loop started with this prompt:
Be an expert in UX and Human-AI interaction research as well as my my mentor. Review the Framer site we have been working on… Help me design a skill, agent, loop or otherwise that I could run on each of them to improve and optimize
Its first pass found that every embedded image on all six case studies collapsed to zero height in any browser window narrower than 1215px, which covers most windows that aren’t maximized. It traced the cause to one breakpoint setting and staged a template-wide fix.
My call: Ship the fix. The extra nav link it suggested stays my decision, and I haven’t made it yet.
The correction that runs the other way
July 4, 2026
Day one, the Archivist guessed my folder of beer-can photographs was brand work for my old deli. My actual reply:
The cans look like this, its just going to be a gallery of an art project I did.
It corrected the record, filed the photography as its own project, and kept the wrong guess visible in the page instead of deleting it.
My call: Mistakes stay in the log. That’s what makes the rest of it believable.
The vault, mapped
The wiki behind this site: every node is a page, every line is a citation or cross-reference. The six case studies you can read here are nodes on this map.
27 pages · 117 cross-links · as of July 2026
Where it gets things wrong
A system that only reported its wins would be marketing. Three documented failure modes, from its own log:
It converged on a mold. The first six case-study drafts came out wearing the same skeleton: identical section headers, identical closers, always exactly three findings. I caught it reading my own site in sequence. There’s now a required check that compares every new draft against its neighbors before it files.
Its flags go stale. It once flagged a hero image as “likely the wrong screenshot.” A later pass opened the file: the image was right and the flag was the bug. Audits now re-verify old flags instead of trusting them.
It guesses wrong at first sight. See the beer cans, above.
Corrections get logged, never erased. The log is the point.
Built the same way
This page is the system describing itself, and it went through the same pipeline: its own wiki page, its own review pass, its own blocker list, and no publish until I said go.
The whole vault started on July 4th, 2026, with me typing I am a complete beginner and know nothing about obsidian. Five days later, the site you’re reading was live.
If this page earns its keep, two things come next: a clickable version of the graph, and a small “last verified against source” stamp on every case study.
colophon · last updated July 2026