the reason your AI-built MVP is garbage isn’t the AI

another week, another client MVP shipped. been doing this for a couple months and here’s what i’ve learned:

write your plan down in docs. be specific - features, flows, constraints. keeps AI focused and stops it from drifting or second-guessing your decisions.
break it into phases. each one well defined before you prompt anything.
one phase per chat. respect the context window. only feed what that phase actually needs.
keep everything in persistent files. specs, decisions, codebase state - outside any single chat. start each new session from those files.
track your progress. what’s done, what’s left, why you made certain calls. otherwise AI will build conflicting stuff across phases.
verify the output. docs with expected behavior + something like playwright to test the real UI. formal tests are optional, some kind of verification loop isn’t.
use work trees to parallelize. run phases in parallel across separate chats, resolve conflicts when merging. this is where the speed really kicks in.

every step compounds. when they’re all in place AI just lands things first pass

1 comment