AI Workflow × Website Maintenance

Multiple AIs modify the same website, why don’t they fight?

When using AI to maintain a website, sooner or later you will encounter version confusion: this task is changed here, that task is changed there, and they cover each other up. It is only after deployment that something is missing. This article uses my own car rollover scene to talk about a set of three layers of protection that you can install by giving it to AI.

Released 2026-07-06 | Last updated 2026-07-06

What is this article talking about?

Launching the website is just the beginning, you will keep changing it after that. When the person making the change becomes several AI assistants, version confusion changes from "occasional" to "inevitable." This article explains three sources of confusion and provides a set of three layers of protection called "release gates": work area isolation, pre-launch inspection, and coercion. The method is packaged into a public skill package. You don’t need to be able to write programs, just leave it to your AI to take care of it. Who is

suitable for?

· People who already have a website and often ask AI to help with revisions
· People who open several AI tasks at the same time and start to think that the file is "weird"
· People who discovered after deployment that the link was broken and the index could not keep up.

What can you take with you?

· Understand the three sources of version confusion and why writing rules is useless
· A three-layer protection approach: isolation, gates, and coercion
· Publish-gate, a public skill package that can be used directly, can be installed by AI in five steps

Let’s talk about a real car rollover scene first

On July 5, 2026, I am revising the official website. An AI was helping me adjust the site navigation. Halfway through, I found that the article data file had been changed, and there was a new course page.

The person who changed it is another AI. It is performing another task, which has nothing to do with this revision, and it is started directly in the same folder. Neither side is archived, and the changes are all mixed together.

What’s the most ironic thing? I wrote this rule in black and white on my work board two weeks ago: "Check the status of a file before moving it." The rules are there, but they are still being violated.

The point of this matterIf you also start using AI to help you maintain your website, sooner or later it will be your turn. It’s not that AI is bad, it’s that “relying on consciously following rules” cannot prevent negligence. So do people.

Three sources of version confusion

1. Multiple tasks share the same work folder

This task is changed to page A, and that task is changed to file B. Each one does not know the other's existence. Git branches cannot solve this problem, because branches only protect the history, not the "site": A has not archived it yet, and as soon as B switches branches, A's semi-finished product will be destroyed.

2. There are too many places that need to be synchronized online, so it depends on human memory

Every time a new article is published on my website, six places need to be synchronized: the article page itself, the article list data file, the sitemap, the site index, the llms.txt read by the AI, and the two-way links between articles. There was an article that missed the process, and ended up suffering from three diseases at the same time: the Internet script was not installed, the English version was not built, and the navigation link was broken.

3. Problems can only be investigated afterwards

The Kanban board records where each task goes, but it is a schedule and cannot control the version. When you find that something is broken, you can only check the records and compare it to find out the cause, but you can't prevent it from happening next time.

Solution: Turn inspection into gate

I later condensed the entire approach into three layers and named it "Publish Gate". The core idea is: leave the inspection to the machine. If the inspection fails, the website will not be able to exit.

First floorOne task and one workshop

Use git worktree to open a separate folder for each task, and only return to the main line after completion. No one is ever working on the main folder. The two AIs are not in the same room at all and cannot step on each other.

Second floorThere is only one entrance to go online

The things that should be checked before going online are written as preflight scripts: whether the link is broken, whether the new page is indexed, and whether the key is written into the public file. Then string it together into a publish script, and finish it all with just "publish".

Third floorYou will also be blocked if you forget to use the gate.

Add a git pre-push hook: If anyone or any AI wants to push the mainline, it will automatically run a check first, but it will not be pushed out. This lock is tied to the repo itself, and no AI from any company can circumvent it.

The first level of instructions looks like this, as long as your AI can understand it:

# One task and one independent workroom git worktree add ../my-site-worktrees/task-name -b task-branch origin/main

The first actual test after installation: the release script was run through one-stop process from inspection to deployment to online verification, with zero manual effort. Later, I deliberately left a bad link to push, and it was really blocked. That moment and the sense of peace of mind that "the rules are written in the document" are completely different.

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.

It does not guarantee: Someone "intentionally" detoured. To seal the detour, the end result is to hand over the deployment to the Git platform for integration (for example, Vercel and GitHub only accept versions that are merged into the mainline). I haven't reached that point myself yet. The current three-tiered company is enough for a one-person company plus a few AI assistants.

Let’s talk about it firstI didn’t invent anything with this set of ideas. The engineering community has been doing CI/CD for decades, and the practices of large companies are much deeper than this. What I did was to reduce it to a size that "people who don't write programs can install it by handing it over to AI."

Want to install it on my website

The entire set of practices is packaged into a public skill package publish-gate, which contains method instructions and four script templates (pre-launch check, one-click publishing, push gate, link scanning), and is authorized by MIT.

publish-gate · Publish gate

Give this skill package to your AI assistant (Claude Code, Codex, or Cursor will all work) and say "Follow publish-gate to install the publishing gate for my website." It will be installed in five steps, and in the end it will deliberately break a link. It has been tested that the gate really blocks.

GitHub Full Description ↗ Look at this package on the skill package download page

Websites grow. It will keep changing, and there will always be version problems. Instead of having to fix it every time something breaks, install the gate and let the machine remember what to check for you.

AI WorkflowWorkflowTool OperationBasicsTeaching Articles