How to Keep Multiple AIs Editing the Same Website From Clashing: The Publish Gate｜Coach Jiang

What this article is about

Putting a website live is only the beginning; after that, you keep editing it. Once the editors become several AI assistants, version chaos goes from "occasional" to "inevitable." This article spells out the three sources of that chaos and gives a three-layer safeguard called the "Publish Gate": workspace isolation, a pre-launch check, and enforcement. The method is packaged as a public skill package, so you don't need to know how to code; just hand it to your AI to set up.

Who this is for

· People who already have a website and often ask AI to help revise it
· People running several AI tasks at once who are starting to feel their files look "off"
· People who only discovered after deploying that links were broken and the index hadn't kept up

What you can take away

· An understanding of the three sources of version chaos, and why writing rules doesn't work
· A three-layer safeguard: isolation, gate, and enforcement
· The ready-to-use public skill package publish-gate, which your AI installs in five steps

First, a real screw-up of mine

On July 5, 2026, I was revising the official site. One AI was helping me adjust the site-wide navigation, and halfway through, I noticed the article data file had been changed and an extra new course page had appeared.

The one that changed it was a different AI. It was running a separate task, completely unrelated to this revision, and it had gone to work right there in the same folder. Neither side had committed, and the changes were all tangled together.

The most ironic part? The rule "check the status before touching a file" was something I'd put in black and white on my work board two weeks earlier. The rule was there, and it got broken anyway.

The point hereIf you're starting to use AI to maintain your website, this will be your turn sooner or later. It's not that AI misbehaves; a rule that relies on good intentions was never going to prevent a slip. The same is true of people.

Three sources of version confusion

1. Multiple tasks share the same working folder

One task edits page A, another edits file B, and neither knows the other exists. Git branches don't solve this, because a branch only protects history, not the live working directory: A hasn't been saved yet, and the moment B switches branches, A's half-finished work is destroyed.

2. Too many places to sync at release, all tracked by memory

Every time my site publishes a new article, six places have to be synced: the article page itself, the article-list data file, the sitemap, the on-site index, the llms.txt that AI reads, and the two-way links between articles. One article skipped the process and ended up with three problems at once: the interlinking script wasn't installed, the English version wasn't built, and the navigation link was broken.

3. You can only investigate problems after the fact

The board records where each task is, but it's a schedule, not version control. By the time you notice something's broken, all you can do is dig through the records and compare; you can find the cause, but you can't prevent the next one.

The fix: turn checks into a gate

I eventually condensed the whole approach into three layers and named it the "Publish Gate." The core idea in one line: hand the checks that matter to the machine, and if a check fails, nothing ships.

Layer 1One task, one workspace

Use git worktree to give each task its own folder, merged back to the main line only when it's done. No one ever works in the main folder, so the two AIs aren't even in the same room and can't step on each other.

Layer 2Only one door to go live

Write the pre-launch checks into a preflight script: whether any link is broken, whether new pages made it into the index, whether a secret key got written into a public file. Then chain it into a single publish script, so one word, "publish," runs the whole thing.

Layer 3Forget the gate and you're still blocked

Add a git pre-push hook: anyone, any AI, that tries to push to the main line runs the checks automatically first, and if they don't pass, the push is refused. This lock is tied to the repo itself, so no AI from any vendor can get around it.

The layer-one command looks like this; it just needs to be something your AI can follow:

# One independent workspace per task
git worktree add ../my-site-worktrees/task-name -b task-branch origin/main

The first real test after setting it up: the publish script ran end to end, from checks to deployment to live verification, with zero manual steps. Later I deliberately left a broken link and tried to push it, and it really did get blocked. That moment felt nothing like the false comfort of "the rule is written down in a doc."

Honestly speaking, what does it guarantee and what does it not guarantee?

It guarantees that every version that goes out through the normal path has passed the check you defined; each version has a version mark, and can be rolled back if an error occurs.

What it doesn't guarantee: someone deliberately going around it. To seal off that route too, the endgame is to hand deployment to a Git-platform integration (for example, Vercel connected to GitHub, which only recognizes versions merged into the main line). I haven't gone that far myself; for the scale of a one-person company plus a few AI assistants, the current three layers are already enough.

A note upfrontThere's nothing I invented here. Engineering has been doing CI/CD for decades, and big companies go far deeper than this. What I did was shrink it down to a size where "someone who doesn't code can hand it to their AI and have it set up."

Want to set it up on your own website

The whole approach is packaged as a public skill package, publish-gate, containing the method write-up and four script templates (pre-launch check, one-click publish, push gate, link scan), under an MIT license.

publish-gate · Publish Gate

Hand this skill package to your AI assistant (Claude Code, Codex, or Cursor all work) and say, "Follow publish-gate and set up a publish gate for my website." It installs in five steps, and at the end it deliberately breaks a link to prove the gate really blocks it.

Full details on GitHub ↗ See this package on the Skills page

A website is something that grows. It keeps changing, so version problems keep coming. Rather than fixing each incident after it happens, put the gate in place and let the machine remember the checks for you.

AI workflowWorkflowTool operationFundamentalsTutorial