The cheapest feature to build is the one users ask for.
The most expensive is the one you assumed they wanted.
The most valuable is the boring one you almost didn't ship.
My model started giving the same answer to every question.
Two days of evals, temperature tweaks, prompt surgery.
The bug: I was caching on a key that ignored the input.
Every question was, technically, the same question.
Watching big tech ship 'AI agents' that need a 40-slide deck to explain, while indie hackers ship a button that does one thing and charge $9 for it.
The demo that needs a narrator isn't a product yet.
How I actually debug a prompt:
1. Swear it's the model
2. Print the exact string I sent
3. Find the typo / stray newline / wrong variable
4. Quietly fix it
5. Tell no one
The compiler can't be wrong when the compiler is vibes.