Lab · 2

If AI replaced the CMS, what stops the site from drifting?

Experiment 1 found that AI can replace a content management system (CMS). But a CMS quietly does something else too: it keeps a site consistent, page after page, without anyone having to think about it. Once AI took over and the CMS was gone, Dr. Paul tested whether a written design rulebook, plus a simple automatic checklist, could do that job instead.

Supported Updated Aug. 29, 2026

Hypothesis

A design rulebook, backed by a simple automatic check, can keep an AI-built website consistent — doing the job a CMS normally does, without a CMS in place.

Finding

Some of the job, yes. Small style details — colors, spacing, fonts — update everywhere at once, because every page pulls them from one shared file, so that part fixes itself. But the CMS's other job, catching bigger mistakes before they ship, doesn't happen on its own. Someone, or something, still has to notice those on purpose.

Some mistakes were caught on this site — not a controlled study, and not enough to prove this works everywhere. But it worked to an extent.

Method

Traditional
Design rulebookCMS enforces the rulesTeam reviewLaunch
Experiment
Design rulebookAutomatic checklist (run by hand)Human + AI reviewLaunch

Launch

A new page or change — often drafted by AI — goes live. Nothing else automatically stops a mistake from shipping.

Drift

Something on the site quietly stops matching the rulebook — the wrong color, a leftover sentence, an inconsistent footer.

Check

Before anything goes live, an automatic checklist scans every page for exactly this kind of mistake — the job the CMS used to do.

Fix

A human or AI reads the results and fixes the problem by hand — the checklist finds mistakes, it doesn't fix them.

Tech

Claudescripts/audit-site.pyShared style fileGitHub Pages

The experiment

The artifact isn’t a writeup — it’s the automatic checklist itself (scripts/audit-site.py), watching over the same AI-built site from Experiment 1, plus the two real mistakes it’s already helped catch, described in the design system’s own rules.

AI handled

  • Writing the checklist
  • Spotting the mistakes
  • Drafting the fixes

Dr. Paul handled

  • Deciding what counts as a mistake
  • Approving the fixes
  • Final review

9 pages checked · 2 real mistakes caught · a clean re-check the day this page went live

Worked: The automatic checklist reliably catches the bigger mistakes — it's flagged nothing since, and stayed clean even while this very page was being built.

Constraint: It can't catch mistakes in the writing itself — repeated wording, outdated phrasing, a wrong reference. That needed a separate tool built just for text.

Finding

Replacing the CMS didn’t remove the job.
It just left it unassigned.

Small style details fix themselves automatically — colors, spacing, and fonts update everywhere at once, because every page pulls them from the same shared style file.

Bigger mistakes aren’t automatic — a leftover sentence sat in one page’s footer until a person noticed it. That’s exactly why an automatic checklist now runs before every page goes live, instead of relying on memory.

Implication

Swapping AI in for a CMS doesn’t just remove a tool — it removes whatever that tool was quietly enforcing. Something has to pick that job back up, on purpose.

If you replace the CMS, replace what it was enforcing too. A design rulebook only works if something keeps checking it against what actually went live.

Next Question · 3

Should Dr. Paul become an avatar?

Not this experiment's question, but the next one in the queue, raised back in Experiment 1: talking avatar, text chatbot, or shelved entirely.

Back to the Lab