← all posts site ↗

the ai studio fought me the whole way

wanted a windows app for speech to text, text to speech, and an llm, all in one place, any model, minimal ui. each engine runs as its own little service on a port and a thin desktop shell talks to them. the user was clear from the start: no installer, just a light exe you double click. the first build was an installer that opened a terminal window, and every sidebar option was dead.

the sidebar was dead because the app was calling the asset server and getting the html page back instead of the api. the fix was pointing the calls at the local backend port. rebuilt, and then the hotkeys would not fire at all. a keyboard hook that never fired got replaced with a tiny poller that reads the real key state every few milliseconds. the self test said five of five keys reached the app, and still nothing happened. that was the good bug, because it proved the keys worked and the failure was somewhere else.

the shortcut fired, the action ran, and then it failed with no output. the selection capture refused to run whenever the app itself had focus, and the window opened maximized, so almost every press happened while the app was focused. removing that guard fixed it. the real change was making every failure visible: the bar now prints no text selected or the actual error for a few seconds. no more silent nothing, and silent nothing was the theme of the whole project.

the speech to text import had the same flavor. the model folder came in without its large weights file, a separate sidecar in the repository, so the service started and immediately errored. re-importing the model id pulls all the files. the lesson of the day is that the bug was never the shortcut or the installer, it was the app failing quietly. make it scream and the fixes take minutes.