How this started
This site didn't start as a home for animated explainers. It started as an experiment: could software written entirely by Claude go to production with no human reading the code, if something checked the results instead?
So the site ran that experiment on itself. Claude maintained it and pushed straight to production with no code review. The contract Claude worked under was published word for word. A nightly job checked the live site and recorded what it found, and a bar at the top of every page reported only what that job had recorded.
The pages below are that experiment, kept as it was. The checklist is the argument in one page, and Legacy Systems is where it stops applying.
Then the videos took over. The explainers turned out to be the part people came for, so the site became about them. This page is how it started.
Vibe coding has a reputation problem. Security vendors write about it to sell scanners. Non-technical founders write about it to sell courses. Almost nobody writing about it shows you a production system they actually run — with real users, a real database, and real consequences if it breaks.
I have. Multiple times. This site is what I actually check, and what I've learned Claude will never check for me.
The proof it started from
First Utah game of the season. I'm testing the live scoring app for my football pick'em product — mine, built by me, with real users on it. Mid-game, from my phone, I push a full UI redesign and wire up a live ESPN data feed — with zero direct access from that phone to my production database.
Here's how that's structured, and why the structure is the point.
The site that checked itself
During the experiment this site was managed by Claude: no PR, no human review of the code. What shipped instead was a nightly process that checked the results, not the code. That was the whole philosophy, applied to itself.
A status bar at the top of every page reported what that process could actually prove: the last run, when it happened, and whether it passed. The nightly schedule was never armed, so every run it recorded was triggered by hand. The diagram below shows the last one.
- 1content / output validationmarkup, leaked secrets, dead links, foreign scriptsok
- 2dependency & secrets scanlockfile vulns, committed env filesok
- 3automated pen test passZAP baseline, exposed routes, headers, TLSok
- 4decisionerrored > failed > passed
- 5rollbackrestore tree to last known-good, prove it, then alert
The original pages
- Checklist →
From the origin story: what AI coding tools covered by default, what they covered only if asked, and what they never covered.
- Legacy Systems →
From the origin story: Claude-to-prod was a greenfield playbook. Here is where it stopped applying.
- Living Checklist →
From the origin story: how a nightly job checked this site in place of code review, and what it recorded.
- My Setup →
From the origin story: the operating contract Claude worked under on this site during the experiment.
- Watch It Get Built →
From the origin story: videos and demos on the method behind software written by AI.