Being a startup founder comes with the constant risk of burnout. Over the last 10 years, I've been hospitalized from exhaustion three times. I don't recommend working 16 to 20 hours a day, seven days a week, for months and years the way I have. There is now no guesswork involved in the assessment: I shaved many useful years of life away working this way. I will die earlier and have fewer healthy years than I would have had I found another way to meet the demands and opportunities. On this side of it, I do regret much of the sacrifice.
But that lamenting is for another time. What I want to talk about today is how AI coding knocked me out again.
One More Feature
I didn't end up at the hospital this time, but I did spend a week on bedrest after a visit to Lake Tekapo (pictured). I had an opportunity (as I've explained in previous posts about building WitFoo Analytics with Claude and migrating our legacy code) to use Claude to rebuild 10 years of work in what turned out to be an eight-month megasprint. I'm very proud of the work. I also recognize how precious and wonderful an opportunity it was to "do it all over again." That opportunity never comes. I really did have a great time in the process.
But the new process has a whole new class of hazards that I did not protect against.
I knew to make sure tests existed for performance, security, and supportability. I created unit tests, system tests, static code analysis, vulnerability fuzzing, end-to-end testing, and a series of soak and burn-in tests to keep Claude honest. I wrote about this in detail in my post on OODA loops and feedback cycles. I built guardrails for the code. I built guardrails for the architecture. I built guardrails for security.
What I did not build was guardrails for my health.
The Dopamine Pipeline
The dopamine release in delivering a feature is as real as beating a boss in a video game, getting a real-life promotion, or winning a high-stakes poker hand. I'm no medical professional, but I am an addict, and I know what it feels like to need one more fix. I didn't realize it until I was laying in bed for the fourth straight day trying to reclaim my health that I had been on my most epic bender. (And I have had some benders that songs have been written about.)
With Claude, I go through a process of designing and planning, then watching it work. For several months, I kept my laptop next to my recliner so I could answer questions, give corrections, or grant permissions. Every time I saw a new feature working, I was overjoyed. That euphoria drove me to do "just one more" feature.
Here's the thing. When I wrote about the value-to-cost ratio shifting, I was looking at the economics. I should have also been looking at the biology. The v:c ratio of dopamine hits to physical cost was also shifting, and I was too busy riding the high to notice. Every completed PR was a pull on the slot machine, and the machine was paying out constantly. That's not a recipe for moderation. That's a recipe for losing track of time, meals, sleep, and eventually your ability to stand up without the room tilting.
A Junkie by Any Other Name
It's been well said that Claude can do 90% of the work better than a good engineer, but it still needs human guidance on the other 10%. That ultimately means I can do 10 times more work. Looking back on the process, many things looked like a junkie operating at peak function.
I would have three to four desktops running one to four Claude sessions each that I was coordinating simultaneously. In my post about the Wizard, Warrior, and Poet framework, I wrote about how the Wizard devises strategy while the Warrior executes. What I didn't mention is what happens when the Wizard discovers he can command a dozen Warriors at once. The Wizard stops sleeping. The Wizard stops eating. The Wizard becomes convinced that if he just stays up a few more hours, the entire kingdom will be built by morning.
It was exhilarating. It was also exhausting. And the gap between those two words is where the danger lives.
Our Drugs Are Getting Better
I've often remarked that I don't think our world is getting better. I think people are more traumatized and have more suffering than we've seen in several generations. I just think our drugs have become more effective.
Good alcohol is cheap and plentiful. We have a video library of distractions in our pockets. Sex in virtually every form is readily available. Video games are insanely detailed and engrossing. We can have candy THC. Our doctors are happy to prescribe us whatever gets us out the door. We can get instant gratification online with engagement on posts like this one.
I wrote in my OODA loop piece about how positive feedback loops are morally neutral. They amplify whatever signal you feed them. That applies to the neurochemistry, too. The same feedback loop that made me a prolific builder was also the feedback loop that ran my body into the ground. The amplifier doesn't judge. It just amplifies.
AI-assisted development is, for people wired like me, the most potent productivity drug ever created. And like every potent drug, the dose that makes it medicine and the dose that makes it poison are closer together than you'd think.
A Powerful Tool. A Powerful Drug.
While I still believe Claude is a powerful tool first and foremost, I also have to confess: to people like me, it's also a powerful drug. It's one more thing to keep me from my thoughts, fears, and traumas. It's one more "good thing" that keeps me from the greatest parts of life (loved ones, nature, being bathed in wonder).
I've written before about sitting in the wonder. About Hagley Oval at sunset. About the view from the Port Hills. About the sound of Mai laughing in the kitchen. Those moments didn't happen during the megasprint. They couldn't. I was too busy watching progress bars and reviewing pull requests and feeling that little jolt of electricity every time Claude said "all tests passing."
I traded wonder for velocity. I'm not sure that's a trade I'd make again.
The Guardrail I Forgot to Build
I spent months building the WitFoo Way, our testing pyramid, our documentation standards, our security scanning pipeline. I wrote about all of it in exhaustive detail across multiple blog posts. I was meticulous about building guardrails for the code. I was meticulous about building guardrails for the process. I was completely negligent about building guardrails for myself.
If I had applied even 10% of the rigor I put into linter.sh to my own daily routines, I'd have caught this. A daily standup with myself: Did you eat three meals? Did you sleep seven hours? Did you go outside? Did you talk to Mai about something other than WitFoo? Did you stop working before 10pm?
I didn't build that script. I should have. I would have caught the regression immediately.
Coming Back Online
Now that my body has recovered (thanks to detox and rest), I need to approach this tool with the caution an addict gives to being at a party. I need to remind myself to go outside. I need to limit my use of the tool. I need to make time to laugh with loved ones.
It was amazing how quickly a new addiction snuck up on me. I've been sober from alcohol for quite a while. I know the playbook. I know how to watch for the patterns. And I still missed it, because this particular drug disguises itself as productivity. It wears a lab coat instead of a leather jacket. It says "let's ship one more feature" instead of "let's have one more drink." The destination is the same.
Wrap Up
If you're a builder using AI-assisted development (and based on the trajectory I described in my value-to-cost post, you should be), please learn from my mistake. Build the guardrails for yourself, not just for the code.
Set a timer. Take a walk. Eat a meal that isn't consumed one-handed while reviewing a diff. Talk to the people who love you about something that has nothing to do with shipping software. Close the laptop when the sun goes down (at least a few times a week).
The OODA loop I described in my last technical post needs one more principle in the orientation phase: the operator's health. You are not a GPU. You cannot be swapped out when you burn out. You depreciate faster than any hardware Anthropic is running, and unlike their clusters, nobody is building a replacement.
I am proud of what we built during the megasprint. I also know the cost, and I don't just mean the tokens. If the Poet's job is to ask whether our actions align with a meaningful trajectory, then I owe the Poet a long overdue answer. The trajectory was fast. It was not always meaningful. And meaningless velocity, no matter how exhilarating, is just another way of running from something.
Go outside. The wonder is still there.
