Death by a Thousand Prompts

By Derek Larson
April 2026

A couple months ago, I decided to reverse engineer Notion, pixel for pixel, in an evening. Just a couple prompts while I cook dinner.

Over the span of eight hours, I maxed out my quotas for Claude and Codex multiple times. Coaxing the agents to burn millions of tokens, as I sent incoherent messages in half a dozen chat threads to keep the plates spinning.

At 4am, strung out, with a migraine coming on, I called it quits.

I don’t use Notion. I’m not trying to build a Notion competitor. But the tokens were there to be spent.

The fatal flaw of chatbots is every message gets a response no matter how stupid it is. They validate every thought, and act on every impulse.

While it’s hard to delude yourself TikTok is useful, but chatting with AI feels productive, like you’re a wizard exerting agency over the world.

Speeding

Over the past 4 months, as a solo founder, I’ve shipped 2 native mobile apps, 2 Claude plugins, 4 open source libraries, and helped hundreds of people delegate work to AI.

I was spending most of my waking hours in conversation with Claude Code, babysitting half a dozen AIs at once. This is framed as a virtue, not a problem.

Daniel Motles on vibe coding

The more I chatted with agents, the more they shrank my worldview, pushing me towards projects AI can handle, not what’s most worth working on.

Loops

Chat is infinite scroll for work.

Just as social media removed the stopping point from entertainment, chat removes the stopping point from productivity.

If AI writes tens of thousands of lines of code I don’t understand, I need AI to help me debug it, which often introduces more bugs, creating a game of whack-a-mole.

Or worse, eroding my ability to make my own decisions, without feeling paralyzed.

Twilight Zone devil tweet

The models improving will only accelerate the problem; compounding the screwups, and encouraging more blind trust.

Break

In 2024, I gave up using most of the internet for 5 months.

Throughout the challenge, I jotted down every question I wanted to research when the challenge was over.

At first, the questions about politics or current events gushed out of me, then turned into a trickle, and then went dry altogether. Consumption was creating the desire to consume.

At the end of the 5 months, I looked back at my list, and not a single question felt worth researching.

Last summer, I decided to go through my last 50 conversations with chatbots, and realized that despite feeling important, none of them moved the needle.

I went cold turkey, and deleted all my AI tools except Claude Code.

At first, this was alarming, like I was losing a limb.

But after a few days, it felt like an AI-induced fog was clearing. I felt less paralysis about decisions, and less time getting caught in AI’s train of thought.

Async

If you’re a manager and every 5 minutes an employee came to you with a question, your brain would explode from the context switching.

Instead, we give humans tasks, they go off and work on them without us. The better the employee, the more time they can work unsupervised.

AI has gotten good enough to do hours of work without us, managing a dizzying number of projects in one thread, making asynchronous delegation realistic.

Over the past two months, I’ve recorded 788 voicemails for my agents. Every half hour Claude Code ingests the voicemail and starts taking action. Fixing bugs in my code, compiling ideas for essays, kicking off research tasks.

Then, a couple times a day I’ll check on Claude’s work, record more voice memos, and go on with my day.

Over 90% of my Claude Code sessions happen without me.

I want someone to solve my problems, not distract me with conversations.

Landing

Being an AI babysitter is easy; there’s no rejection and no failures a prompt can’t fix.

But shipping a product worth using is hard. Writing an essay you’re proud of is hard. Closing sales is hard. Maintaining relationships is hard.

AI should be in the service of us, not us in the service of AI.

Subscribe for updates