Building software I actually use with AI.
AI writes most of the code. My job is figuring out what actually needs to be built, designing the workflow, testing the result, and improving the tools through real use.
Task 4 was scoped as a one-line safety check — verify a project wasn't already registered before touching it. It turned into a 402-file backfill script, because the old projects weren't duplicates to catch. They simply weren't in the database yet.
What happens when the same project is saved from two places at once. I designed a two-step reservation system to stop the collision, then realized I wasn't preventing the race — I was rebuilding something the database already did.
Once identity was solved, I assumed the hard part was over. I treated saving and registering as one step — until I realized a browser download and a network call don't fail the same way.
I wanted a simple rule: if a project's already registered, don't register it again. Using the filename as the key seemed obvious — until two unrelated projects collided under the same name.
I kept trying to identify things by their names — filenames, titles, playlist labels — and each time it broke the same way. Here's the pattern I didn't notice until the third time.
Two tools needed to share data. Merging them felt obvious — until I noticed the real problem wasn't project boundaries, it was scrambled ownership of a few specific pieces of data.
I wanted my video editor to look more dynamic. Five layers of "why" later, the real problem wasn't the UI at all — it was a hand-edited JSON file I kept breaking.
I don't write code, but I build most of my own tools with AI. This blog documents that process — and the real lesson so far: the hard part was never writing the solution, it was figuring out what the actual problem was.