What does this article talk about?
A webpage you got halfway through on your laptop, picked back up on your desktop, and the files are either stale or broken; and the one you finished on your desktop, you can't even find when you try to edit it on your laptop while out. If you're like me, running two AI assistants, Claude Code and Codex, on each of two computers, that's four environments all touching the same project, and this is almost bound to happen. This article records how I moved a webpage project out of iCloud and into git in a single evening, so all four environments share one source of truth. You don't have to type any git commands yourself throughout the migration; you just give the AI the right instructions and let it do the work.
- People with two or more computers, a laptop when out and a desktop at home, who sync files through iCloud or Google Drive
- People already using AI assistants like Claude Code or Codex to build webpages or small projects, who are starting to hit the "but I fixed this yesterday" problem
- Knowledge workers who don't code, have heard of git, but have always thought of it as an engineer's thing
- A rule of thumb: what belongs on iCloud, and what must go into git
- A migration workflow you can follow step by step, including how to verify it
- A three-layer approach to "where to write a rule so every AI can see it," useful in any multi-AI setup
Short on time? Jump straight to "How you can start" and grab the prompts first.
Real-world scenario
When I build webpages, Codex and Claude Code take turns on my desktop, and I have the same pair set up on my laptop. On the night of July 11, 2026, I raised a long-standing frustration with the Claude Code on my desktop: when I build a webpage on my laptop while I'm out and then want to pick it up on my desktop back home, it often goes wrong. The reverse happens too: a webpage I finished on my desktop, I try to tweak a small thing on my laptop while out, and that goes wrong as well.
Taking stock, my working setup looked like this: one desktop and one laptop, each running Claude Code and Codex, four AI environments in total. The project files lived in my iCloud knowledge base, on the assumption that iCloud would sync automatically and both computers would see them.
Sounds reasonable, right? I thought so too, until something telling happened that night.
For a document, arriving two minutes late is fine; you just wait a moment. For code, in those two minutes I'd have kept editing on the laptop, working on the old version, and the two sides fork from there. After that, whoever overwrites whom, the work is wasted.
Problem Source
Let's break it down. The issue lies in three places.
First, iCloud synchronization is designed for files.
iCloud has three features that work well with files but cause problems when used with code:
That two-minute demo was evidence. Files can wait, but code cannot; you don’t know if you’re editing the latest version.
Two computers both edited the same file, and iCloud often creates a "File 2" conflict copy or only keeps one; even if you skip past the prompt, it won’t tell you which version is correct.
To save space, iCloud turns files you haven't opened in a while into cloud placeholders. The folder looks fine to a human, but a tool that expects the file to already be local can fail or time out the moment it hits a placeholder that hasn't finished downloading. Earlier that same day I'd dealt with another case: a scheduling tool kept in iCloud kept timing out because of placeholder files, and again the fix was to move it off iCloud.
Two: four AI environments with disconnected memories
This is a problem specific to multi-AI workers. First, here are the four environments I mean:
| Claude Code | Codex | |
|---|---|---|
| Desktop | Environment ①, with its own memory | Environment ②, with its own memory |
| Laptop | Environment ③, with its own memory | Environment ④, with its own memory |
Four quadrants, four separate memories that are not connected. Claude Code on the desktop has its own memory, while Claude Code on the laptop has another one. Codex is completely unable to read Claude’s memory. I established certain rules with Claude on the desktop; none of the other three environments know about it.
So, even if the move was done correctly, as long as the "where is the new home" information only exists in a single AI's private memory, the other three environments will still follow their old paths to go and update the old files.
Three, old paths will remain embedded in various documents.
In my knowledge base, there are index pages and operation manuals that mention "the source code is located in a certain folder." After the move, these sentences all become landmines. Any AI that reads them will be directed to the wrong place.
The mechanism that fixes it
Core principle: documents go through iCloud, code goes through git
The test is simple. Ask yourself one thing: if a small part of this breaks, would I be able to tell?
Articles, notes, meeting records. They sync a couple of minutes late and occasionally spawn a conflict copy, but a human can see at a glance how to handle it. iCloud is fine for these.
Webpage projects, scripts, config files. One wrong line and the whole thing breaks, and it breaks silently. Git works by reconciling: before you start, you reconcile with the center (pull), and when you finish, you report back (push). If both sides really did edit the same spot, it tells you exactly where the conflict is and lets you choose; and as long as you've committed, both sides' content can be recovered.
Moving Process (the version I personally walked through that night)
List all web projects in their respective locations, identifying which ones are still active and which have become history. I had the AI scan the entire knowledge base to find the active projects residing on iCloud.
Copy the project to your local Developer folder, create a git repository, and compare it with the original files for consistency. Keep the original directory intact; this is your fallback option in case of any issues.
Private means that only you can see the source code; the website was meant to be deployed publicly, and visitors won't be affected at all.
Leave a note file in the original location that clearly states "This folder is frozen. The source of truth is in such-and-such repo; editing here does nothing." Any AI passing through the old home gets pointed to the new one.
Change any sentences in the index page and user manual that mention old locations, replacing them with the correct new location. Ensure no outdated references remain.
Test on a laptop, write a test line, push it up; then pull down from the desktop machine to review and delete the test line before pushing back. Completing this loop confirms that the synchronization link is working smoothly.
That evening's real-time testing completed the entire process in less than an hour, with both machines passing the double-check successfully.
Where should rules be written: Three layers
A key lesson in multi-AI environments: The scope of a rule depends on where it is written.
All AI will read this master file first, which contains the rule "Code should be git and stored locally by developers." This rule applies to all four environments.
A rule that travels with the file, written into the old folder. No matter which AI reaches the old file, or by what path, the warning is the first thing it sees.
What AI remembers is only useful to itself. Convenient, but make sure you don’t store critical information like “Where’s the Truth?” in just this layer.
The closer a rule sits to the file itself, the harder it is to miss. My order is: write the signpost note first, then update the global rules, and treat private memory as nothing more than a cache. Here is how to do each of the three layers:
Routine after that: One Rule
What it guarantees and what it doesn't guarantee
What it guarantees: as long as you follow "pull when you start, push when you finish," the silent iCloud accidents, the quiet overwrites and quiet forks, stop happening. Git records every commit, and even when there's a real conflict it lays it out in front of you to choose; anything you've committed can be recovered.
How do you start?
Your first step could be to let your AI help you inventory. Pass this section to your Claude Code or Codex:
After completing the inventory, pick one of the most active projects to migrate first. Let AI assist with this process, moving one project at a time:
After thorough verification, scale up to other projects. After migration, remember three things: label the old directories, add a global rule, and update all documents with old paths.
Recap
- Two computers and four AIs on the same project are bound to clash over iCloud, because its sync was designed for documents.
- The rule of thumb in one line: anything that can break in a small way a human wouldn't notice belongs in git.
- Six steps to move: take inventory, copy without deleting, push to a private repo, leave a signpost, update old paths, and verify across both machines.
- Where you write a rule decides who sees it: the global rules file for all AIs, the signpost note travels with the file, private memory is only a cache.
- One daily rule: pull when you start, push when you finish.
Common Traps (and Countermeasures)
A reminder about where I stand
The problem here is very technical, but the core is the same thing I keep saying: however smart AI gets, it still needs you to set the environment up. Four AI environments stay out of each other's way thanks to one set of rules everyone can read, one single location for the truth, and one rule simple enough that you won't forget it. Get the foundation right and AI can genuinely help you accumulate, with all four helping you at once.
I am Coach Jiang
Tacit knowledge distiller and AI application planner. I hold two free online talks every month, sharing hands-on experience and methods. If these topics interest you, if you want to keep learning, or if you need consulting, you are welcome to start from the community.
Main topics: using AI as a thinking partner to improve decision quality and depth of thought; organizing knowledge and experience into prompts, skill packages, and knowledge bases so AI can use them flexibly.
Interested in AI × Knowledge Management and Tacit Knowledge Extraction? Start here.
Join the Community →