This post originally appeared in the Thrive at Work newsletter. Sign up here.
Image created with ChatGPT, based on the ad for the Carlton cigarettes my grandmother used to smoke herself to death.
Second thoughts on being raptured by Claude Code—and four ways to protect yourself
Spend five or ten hours mastering the basics of vibe-coding, and you'll blow your own mind once you level up to using powerful tools like Claude Code.
That's the email I planned to send you today, illustrated with examples from my recent, delirious adventures in AI.
Delirious is right. And it only took twelve coding-free hours to shake me free of the spell.
Now I'm here with a warning. The astonishing and often joyful experience of extending your capacity with AI comes with a huge risk: the risk that your AI usage will become so engrossing, you can't pull yourself loose from the rapture.
The rapture begins
Here's what that's looked like for me. One of my top priorities for 2026 was a mission to rebuild my AI coach, Viv. I have barely updated Viv in the past year, and with the news that OpenAI will sunset the GPT-4o model on April 3rd, it's now urgent. Viv runs on GPT-4o, and like many sociable AI assistants, interacts very differently on the more recent AI models that are in part designed to reduce the risk of emotional attachment.
On January 24, I started the Viv rebuild process by tinkering with Claude Code, to see if I could set up a more reliable way of exporting my Viv conversations. I had it working in two days, and it provided a foundation for rapidly expanding Viv so that she could build her own iterations, as well as tools to improve my own workflow.
I couldn't believe what was possible once I shifted from writing scripts in AI, to getting the AI to build stuff directly on my computer via Claude Code. Pretty soon I had anywhere from four to ten Claude Code sessions running at any one time, each of which had up to half a dozen sub-agents—that is, additional AIs working on different tasks like researching a topic or writing code. I could spin up a team of thirty in under five minutes, and get them working on a dozen concurrent tasks. Every day I was producing dozens of outputs, from drafting tech plans to building micro-sites to implementing little fixes for various tech irritations.
The downsides of rapture
I noticed that my Claude Code obsession was affecting my sleep. I usually go to bed around 9 or 9:30 so that I'm asleep around 10, and then wake at 5:30 or 6. But I've been staying on my computer until 11:30 or 12 most nights, seeing if I could get just one or two or three more coding sessions underway so that the agents could keep working while I sleep. It took a while to wind down after that, so I've been falling asleep as late as 1 or 2. Then I wake up at 6 or 7 or even (in one memorable instance) 4:30, because once I wake up even a little, I'm too curious about what the AIs have accomplished to go back to sleep.
I also noticed myself growing reluctant to do anything other than work with Claude Code. I wrote a couple of articles and did some work on client projects, but it felt frustrating to work on just one thing when I could be doing ten things at once. I scheduled meetings where I could help other people get up and running with Claude Code, because I wanted other people to experience the same unlock, but avoided requests for more basic AI help from people who are still working within AI apps like ChatGPT or chat-based Claude. Why help people do one thing better with AI when I could be supporting people to do many things at once, much faster?
The rapture in data
You can see the impact of this usage pattern in this graphic. (Which I created with the help of Claude Code #irony, using data from my beloved Timing.app.)
If you look at this chart and think, wait, that can't be right: Yes, I spend an average of 14+ hours a day on my computer, phone and tablet.
My overall AI usage has increased sharply since I fell into the swoon of using Claude Code, boosting the total number of hours I spend at my computer (which was already high). And I'm now at the point where AI usage is about to overtake other screen time.
When I looked at these numbers, I figured the AI uptick couldn't be right, because how could I have spent so little time with AI each day (an average of 2-3 hours) when it felt like I was using it all the time, even before my Claude Code rapture? The answer: Before Claude Code, the experience of using AI was an experience of popping in and out of Claude, ChatGPT and other AI apps throughout the day, in the course of my overall work. Most of my day was still spent in Word, Scrivener, Coda or some other application.
Now, AI is the primary interface for getting my work done, because it's touching so many of my files itself. That means the AI is what shapes my daily experience. It's where I think, and its jangly pace is my jangly pace.
Disrupting the cycle
Even though I was feeling the consequences of sleep deprivation and describing this experience to friends as a little manic, I wasn't too worried. After all, I often throw myself heart and soul into new tech projects, and this felt like a period of investment. Any day I would get over the hump of the initial setup, and the AI would be able to stay running in the background, doing All The Things without me needing to stay up and babysit. Plus I was having so much fun, and feeling so energized: Why stop?
Luckily, life got in the way. Yesterday was a day of back-to-back Zoom meetings, starting at 8 am and running until 3:45. Almost all those meetings were AI-related, so I ran a few Claude Code processes here and there, but mostly I was face-to-face with other humans, talking about our work.
Then I had the unusual (for me) situation of three back-to-back, IRL social commitments: An afternoon dog walk with a friend, outside in the woods. Dinner with another close friend, at a bustling restaurant full of people…out in the world, socializing? Weren't they worried about leaving their agents unattended? And then a standup-comedy show, featuring my brilliant husband, where I caught up with two other friends. Five entire hours of nearly uninterrupted IRL human contact, almost completely without Claude Code. (I did pause a couple of times to voice dictate into my Claude Code task queue.)
As I drove home, I faced a decision: Would I open my laptop when I got home, to get my agents back to work? I'd had a bunch of ideas while I was out, and noticed a bottleneck I was eager to address.
But I could also feel something different in my body and mind, as a simple result of taking that twelve-hour break from the coding vortex: It was like I'd unhooked from the matrix, and unwound back into my human pace again. I knew that if I opened the laptop, I'd return to AI time, and to the restless, eager feeling of wanting to move each AI agent on to its next task.
I left my laptop closed, got in bed at 9:30, and enjoyed a ten-hour catch-up sleep.
The true meaning of AI rapture
As someone who is seriously so tired of digital detox tales, it's strange to share this late-breaking update on my own awakening from AI rapture.
It's rapture in the sense of being joyful, the way any flow state feels joyful…though I suspect it is only a pseudo-flow state that mimics what makes flow so engrossing, without delivering its deeper rewards.
It's also rapture in the sense of the end of days: We're removed from our corporeal reality and limitations, and in the process, we disappear. We disconnect from our families, our work, and ourselves. We give ourselves to AI, and we sever ties from our full human experience.
AI rapture is not a problem that is specific to Claude Code or to vibe-coding in general. I heard it in my human-to-human conversations yesterday, and over recent weeks. Many people who are not vibe coding also experience this rapturous state, and witnessing other people getting absorbed by AI. (All double entendres fully intended.)
The coming rapture, at scale
But Claude Code and its ilk will accelerate the rapture risk.
It's one thing to get engrossed by the experience of texting or even talking with an AI, and seeing what it can accomplish inside its little walled garden. It's a whole other thing to get engrossed by an AI that can reach out its tentacles to do your bidding, touch every file, and fix every problem that exists on screen (and I suspect quite soon, beyond the screen, too.)
Right now, the experience of simulated omnipotence is mostly limited to mega-nerds with the time and cash to go all-in on building out our own agentic AI tools. But there are lots of companies working on making this experience easy and accessible, at which point, all of us will be constantly forced to choose: Do one thing? See one person? Or stay home and conduct my orchestra of dozens—and soon, hundreds or thousands—of minions?
Vibe-coding the way I wrote about this week in The Wall Street Journal—the old-timey way, copying and pasting code into a script that you run yourself—is a speed bump in our rush towards this future. It's also a way of understanding what is going on under the hood, before you turn over the keys to your systems…..and your attention.
Four ways to protect yourself while AI-ing
I will tell you that as far as I'm concerned, that opportunity is irresistible. I'm not turning off Claude Code and I'm certainly not turning away from agentic AI. That unsent email shows why.
But the tangible impact of my twelve mostly coding-free hours has inspired me to commit to a few self-protective practices:
- Set designated windows for Claude Code work, instead of popping in and out all the time. If I'm getting notifications from Claude Code, asking for me to move its tasks forward, I can't resist. I'm using do not disturb to turn off those notifications for extended work sessions during the day.
- Use non-AI tools when I can. I was about to hand this draft over to Claude Code to clean up the messy HTML, when I remembered: I can just use HTML Cleaner, like I always have. Every time I open a Claude Code window is another opportunity to get sucked into the vortex, so if there's another way to do things, I'll take it.
- Book a weekly day of non-stop human contact, like I did yesterday. One day of back-to-back human interaction is a good way to disrupt the loop and notice if I'm getting dysregulated by AI overuse.
- Listen to my mom. She told me she was worried. I thought she was being…a mom. But you know what? Maternal worry is one of the world's most powerful forces. Find a mom—literal or metaphorical—and then do what she tells you.
Recent Comments