The first time you watch an AI assistant edit a page on a live WordPress site, you feel two things at once: this is the future, and this could go horribly wrong. Both feelings are correct. The difference between reckless and routine is not the model — it is the guardrails around it, and they are the least glamorous, most important part of any AI tooling you let near production.
We build Divi Autopilot, which lets Claude build and fix Divi 5 sites, so this question is not theoretical for us. Here is the safety model we ship, presented as a checklist you should apply to any tool that writes to your site — ours included.
1. Every write gets journaled
Not “there is a log somewhere.” A journal: what changed, on which page, in which module, from what to what, when, triggered by what. When a client asks “why does the About page look different?”, the answer should take thirty seconds to find. An AI edit with no paper trail is a rumor with side effects.
2. Undo works per change, not per catastrophe
Site-level restore is the nuclear option — it takes the good changes down with the bad one. The guardrail that actually gets used is per-fix undo: this one alt-text change, this one image compression, reversed individually, leaving everything else standing. If a tool can apply fifty fixes but only undo them as a block, it can really only undo zero.
3. Snapshots before anything risky
Before structural work — migrations, design-system changes, anything touching many pages — take a snapshot you can restore. Cheap insurance, and the discipline matters as much as the feature: the agent should take it automatically, because the human forgets exactly when it matters most.
4. Dry runs by default
Autopilot’s fixers are dry by default — asked to fix something, they first return the plan: here is what would change. Writing requires the explicit go. That one design choice converts “the AI did something” into “the AI proposed, I approved,” which is the relationship you actually want with software that edits your business.
5. Verification that doesn’t trust the AI’s word
After changes, something other than the thing that made them should check the site. That is what the audits are for: re-run, read the diff — fixed, introduced, persisting. “Introduced: 0” is the sentence that lets you sleep. And the plugin’s own self-test — about 1,650 checks — verifies the tooling itself against your install, because guardrails that are never tested are decoration.
6. Scopes and seats that mean something
The connection should be a credential you issued and can revoke, not a shared admin password. Licenses should bind to sites you chose. Boring, yes — boring is the point. Security features are the ones you want to be bored by.
The honest trade of letting AI edit a WordPress site
With these guardrails, AI editing stops being a leap of faith and becomes what it should be: delegation with receipts. You hand over the tedious work — audits, fixes, bulk operations — and keep the judgment, the approvals and the undo button. Without them, you are pasting your site into a slot machine.
See the guardrails in person: connect Claude to a Divi site and run an audit — reading is free, and the journal starts on day one. Plans from $99/year.