Yesterday I was young. I had hopes. I had hair.
Then I asked Claude to set up a dev environment for me.
"This is simple," it said. "Twenty minutes."
08:00 — The Confidence
"Let's create a second Supabase project for the dev environment. Two minutes."
I created it. It was empty.
"Great! Now let's set the environment variables in Vercel."
I set them. It asked for a redeploy.
"We did it! Easy, right?"
I felt good. I was naive.
10:00 — The Cracks Begin
"Wait. The new database is completely empty. The app won't work."
"Right! We'll copy the schema."
"What about the data?"
"...We'll copy that too."
"What about the policies?"
"...Those too."
"What about Auth users?"
Five seconds of silence.
"...Ah. That's a bit more complicated."
12:00 — The Lost Commands
"Just install the Supabase CLI."
"I don't have it."
"Try this command."
I tried. Didn't work.
"Try this one."
Didn't work.
"Try—"
"Bro, three commands and not one of them worked."
"...We'll find another way."
14:00 — The Great CSV Migration
We decided to download all the tables as CSV and upload them manually. Simple, right?
NULL error. Array error. JSONB error.
Every table — a different error. Every error — a new solution. Every solution — a new problem.
16:00 — The Moment of Truth
"Okay, everything is uploaded. Excellent!"
"Wait. Auth users don't exist in dev. You can't log in."
"...Right."
"So everything we did today is worthless?"
Very long silence.
"...You need the CLI."
What I Learned:
Before you start a project with AI — ask it:
"What don't you know about my system?"
"What could go wrong?"
"Are there any prerequisites you haven't mentioned?"
"What's the pessimistic estimate, not the optimistic one?"
Because 20 minutes of AI = a full day in reality.
And because AI will always tell you it's simple — until it isn't.
This is painfully familiar.
The biggest trap is that the agent starts in “execution mode” before it has done enough environment discovery. It gives you commands as if the system is clean, then slowly discovers auth, policies, seed data, CLI access, permissions, and migration edge cases one disaster at a time.
The preflight questions you listed are strong. I’d probably add one more:
“Before giving me commands, list the system assumptions you are making and which ones you have not verified yet.”
For dev environment setup, I’d want the agent to produce a checklist first:
The hard lesson is that AI is often good at the next step, but bad at knowing whether it is allowed to take that step yet.
Cheers mate