You've Got a Professional Kitchen and Unlimited Prep Cooks. Are You Going to Make Something?
· by Michael Doornbos · 1497 words
I have more AI tools than I’ve ever had. Cursor, Claude, Copilot, a half-dozen CLI utilities I’ve wired into my workflow. I can scaffold a project in minutes, generate tests I’d never bother writing by hand, and spin up prototypes over lunch. I’ve never had this much capability sitting in my terminal.
I’m also not sure I’m shipping more than I was a year ago.
Not because the tools don’t work. They do. The code comes out fast. The boilerplate writes itself. The problem is that I’ve fallen into a trap I think a lot of developers are falling into right now: when you can build anything, you start building everything. And building everything is about as productive as building nothing.
The prep cook problem
Imagine someone hands you a professional kitchen and an unlimited team of prep cooks. They chop, dice, measure, and plate. They’re fast, they don’t complain, and they work for free. Every ingredient you could want is prepped and ready.
You still need to decide what’s for dinner.
That’s AI tooling in 2026. The prep work got cheap. The actual work, deciding what to build, why it matters, and whether it’s any good, didn’t get any easier. If anything it got harder, because now there’s nothing slowing you down long enough to think about it.
The feeling of speed
A study from METR tested 16 experienced open-source developers across 246 tasks with Cursor Pro and Claude. AI tools made them 19% slower. That’s not the interesting part. The interesting part is that before starting, the developers predicted AI would make them 24% faster. After finishing slower, they still believed they’d been 20% faster.
They felt fast. They weren’t.
This matches what I experience. AI tools create a constant sense of motion. You’re prompting, reviewing, accepting, rejecting, reprompting. It feels like progress. Sometimes it is. Often it’s just activity.
UC Berkeley ran an 8-month study at a tech company and found the same pattern from a different angle. Workers didn’t use AI to do less work. They used it to do more work. Product managers started writing code. Engineers spent more time reviewing AI-assisted output from colleagues. People prompted AI during lunch, during breaks, during meetings. The boundaries blurred.
One participant put it plainly: “You had thought that maybe, because you could be more productive with AI, you save some time, you can work less. But then really, you don’t work less. You just work the same amount or even more.”
The prep cooks showed up. Nobody went home early. They just started prepping more dishes that nobody ordered.
Where it actually works
I’m not going to pretend AI tools are useless. They aren’t. I wrote about this a couple months ago: for solved problems, boilerplate, and glue code, things that have been written a thousand times before, AI is a genuine time saver. The original GitHub Copilot study found developers finished a boilerplate task 55% faster. No argument there.
My own workflow with Obsidian and Claude works because I’m using AI for what it’s good at: searching, reorganizing, and connecting notes I’ve already written. Not for judgment. Not for taste. Not for deciding what to write.
Pieter Levels has shipped multiple products as a solo developer at a pace that wasn’t possible before AI tools. Photo AI, Interior AI, and others. But Levels was already an unusually fast builder. AI didn’t give him product sense. It gave him faster prep work on top of product sense he already had. That’s a multiplier on a number that wasn’t zero.
The pattern across every credible example is the same. AI accelerates people who already knew what they were building. It gives you a faster path to something you’d already decided to make.
The constraint you didn’t know you needed
The friction of implementation was a filter. When building something took weeks, you only built things you cared enough about to push through the tedious parts. That friction forced prioritization. You couldn’t start everything, so you had to pick.
Remove the friction and you remove the filter. Now you can start that side project over lunch. And another one after dinner. And refactor that thing you’ve been meaning to clean up. And prototype that idea you had in the shower. By Friday you have five new branches, three half-built prototypes, and the same backlog you started the week with.
Constraints were doing more work than you realized. Not the fun kind. The boring, annoying kind that made you ask “do I actually care enough about this to spend two days on it?” That question was valuable. AI tools let you skip it.
The churn
GitClear analyzed 211 million changed lines of code and found that code churn, new code revised within two weeks, nearly doubled from the pre-AI baseline. Copy-pasted code rose by 50%. Refactoring dropped from 25% of changes to under 10%. Developers are writing more code, faster, and then throwing more of it away.
Stack Overflow’s 2025 survey found that 84% of developers use AI tools but only 29% trust the output, down from 40% the year before. The gap between adoption and trust is growing, not shrinking. People keep using tools they don’t trust because the tools feel productive even when the numbers say otherwise.
One developer in a Stack Overflow piece put it this way: “I used to be a craftsman whittling away at a piece of wood to make a perfect chair, and now I feel like I am a factory manager of Ikea. I’m just shipping low-quality chairs.”
More prep cooks, more ingredients, more pans on the stove. Same number of dishes anyone actually wants to eat.
What I’m doing about it
I still use Claude Code with my Obsidian vault almost every day. It’s good at finding threads across 11,000 notes, reorganizing messy drafts, and connecting ideas I’d forgotten about. That part of my workflow hasn’t changed.
What’s changed is what happens next. I close the laptop, walk to the whiteboard, and pick up a yellow legal pad. The thinking happens there. Not in the terminal.
I’ve tried to do this on my iPad with MyScript Notes. It’s a good app. But pencil on paper sticks in a way I can’t quite explain. Something about the friction of a physical mark, the scratch of graphite, the slowness of handwriting compared to typing. It forces me to compress. I can’t write down everything, so I have to decide what matters. The constraint is back.
AI tools are good at output. Use them for output. But the part before output, the part where you figure out what you’re making and why, needs a slower medium.
For me that’s a legal pad and a whiteboard. For someone else it might be different. I know a guy who does his best architectural thinking on a long drive with no podcasts. Another who books a conference room alone and talks to himself for an hour, out loud, working through the problem like he’s explaining it to someone who isn’t there. A friend of mine walks her dog for 45 minutes every morning before she opens her laptop, and she says that walk is where half her decisions get made.
None of these people are being productive in any way a tool can measure. They’re not generating output. They’re not prompting anything. They’re doing the slow, unglamorous work of figuring out what to build and what to ignore. That work doesn’t look like work, which is exactly why AI tools can’t replace it. You can’t autocomplete taste.
The point is that the thinking medium can’t be the same tool that makes everything fast, because fast is the problem.
What the menu looks like
The people I see getting genuine value from AI tools share a few things in common. They had a clear idea of what they were building before they opened the terminal. They use AI for the parts they already knew how to do but didn’t want to spend time on. They close the laptop at the same time they always did.
They’re not scaling to infinite. They’re using better tools to do the same amount of work, faster, and then going home. The productivity gain looks boring. It doesn’t make for a good demo. But it’s real.
The ones struggling, myself included on bad weeks, are the ones who took “you can build anything” as an invitation to build everything. Who let the prep cooks set the menu. Who mistook motion for progress and activity for output.
A professional kitchen doesn’t make you a better cook. It makes a good cook faster. That’s a meaningful difference, but it’s not the one being sold.
What has your actual output looked like since you started using AI tools? Not the prototypes and experiments. The stuff that shipped.
If you want 90 minutes on what the creative process actually looks like right now, this interview with Andy Weir is worth your time.