When Claude Code Went Down, I Stopped Coding

A Holiday, a Stomach Bug, and an Outage

It was early March, a holiday weekend. I'd been looking forward to finally touching a side project I'd been putting off with the excuse of being too busy. I'd set aside the day before to get something done — and then, naturally, I came down with a stomach bug. Spent the whole day in bed, groaning.

By evening I'd recovered enough to sit down for some coding, only to find that Claude Code's servers were completely dead. The auth server, the console, Claude Code itself — all down. According to Anthropic, they'd been dealing with "unprecedented demand" over the past week. Users fleeing ChatGPT over OpenAI's Pentagon contract backlash had flooded into Claude, and the service couldn't hold up.

But the outage itself wasn't the real problem.

The Guy Who Just Stopped

When Claude Code went down, I simply didn't code. It never even crossed my mind to open a terminal and write code by hand. I just thought "eh, I'll do it when the servers come back," and closed my MacBook.

Then I realized something even funnier. I have an OpenAI Codex Pro account. Codex CLI is a terminal-based coding agent very similar to Claude Code. It's open source, included with ChatGPT Plus, and already installed on my Mac. And yet not once — not for a single second — did the thought cross my mind to try Codex.

Isn't that a little strange? A perfectly good alternative was right there beside me, and I couldn't even think of it.

Anchoring, Status Quo Bias, and Tool Lock-in

At first I wondered if this was anchoring. But anchoring is the bias where the first piece of information you encounter sets the baseline for all subsequent judgment. That's not quite what was happening here.

A closer concept is status quo bias. In psychology, status quo bias is the tendency to prefer the current state of affairs — a mix of loss aversion, emotional comfort, and decision avoidance. It wasn't that I failed to think of Codex because Codex is bad. It was because I was cognitively locked into Claude Code as my default state.

When I think about it, this isn't something that only happens with AI coding tools. Vim users don't use VS Code. Chrome users don't use Firefox. But text editors and browsers are habits built over decades. AI coding tools have only existed for a year or two, and they've already produced this level of lock-in. That's what I find interesting.

How Different Are Claude Code and Codex, Really?

Articles comparing the two describe Claude Code as a "collaborative partner" and Codex as a "delegated employee". Claude Code deeply understands context, asks clarifying questions, and improves iteratively. Codex is more autonomous — you hand it a task and it runs with it.

Theoretically, switching between the two depending on the situation is the rational thing to do. If Claude Code is down, switch to Codex. In fact, many developers recommend using both in tandem: Claude Code for precise work, Codex for autonomous execution.

But in reality, I didn't do that. I wasn't rationally selecting the best tool — I just stopped when the familiar one wasn't available. Like how, if your regular café is closed, instead of going to the one next door, you just don't have coffee at all.

The Age of Tool Dependency

Apparently this isn't just me. According to MIT Technology Review, one game infrastructure engineer who tried working on a side project without AI tools said: "things that used to be instinct became manual, sometimes even cumbersome."

Anthropic's own research also turned up an interesting result. An investigation into the impact of AI coding assistance on skill formation found that the group using AI scored 17% lower on proficiency tests compared to the group that didn't. The biggest gap appeared specifically in debugging ability.

My situation is a bit different from skill degradation, of course. It's not that I can't code anymore. It's just that the will to code without AI has evaporated. This isn't a capability problem — it's a behavioral pattern problem. More honestly: "coding without AI just isn't fun anymore."

AI Tools as a Single Point of Failure

In software engineering, a single point of failure is a component that, when it goes down, brings the whole system to a halt. On March 2nd, Claude Code was a single point of failure in my coding workflow.

The Register reported that social media was flooded with reactions like "now that Claude is down, developers have to go back to their pre-AI habits and write code by hand." Self-deprecating humor, sure — but probably half-sincere.

Looking at it coldly, my workflow is fragile. I depend on a single service, and even with alternatives available, I'm not prepared to switch. In an enterprise environment, there would obviously need to be a failover plan for a structure like this. I never expected I'd be thinking about this seriously at the level of an individual developer.

So What Am I Going to Do About It?

Honestly, there's no particular conclusion. I could resolve to set up Codex as a backup, or commit to maintaining coding habits that don't depend on AI — but let's be honest, I probably won't do any of that. When Claude Code comes back up, I'll go back to using it exactly like before.

But one thing did become clear through this experience. I had stopped being a "user" of AI coding tools and had become a "resident." I wasn't choosing the tool; I was living inside it. When the tool stopped, I stopped. And probably, this isn't just my story.

If there's a lesson from the holiday, it's this: a stomach bug clears up in a day. Tool dependency is a lot harder to notice.