Inspiration. AI now writes a lot of minutes and summaries, and most get approved without anyone listening to the recording again. Plenty of tools create content from a recording. I wanted one that checks it.
How I built it. One SKILL.md that coordinates two of the NAS's own experts: one transcribes the recording, the other compares the transcript with the written file. The main model only passes fixed messages between them.
Challenges. No scripts, no device to test on, and a small local model. I tested the orchestration in a simulated runtime with a weak model and rewrote the instructions until it followed them.
What I learned. Writing the Skill took half an hour. Finding out what a small model does with it took the rest of the weekend.
