SkyWatch Command

Mouse, gesture and voice all funnel into one command dispatcher. Every feature is input-agnostic by construction: add a behavior once, all three inputs inherit it. The architecture is the thesis.

The app kept transcribing its own voice, it heard itself say “Done” and tried to run it as a command. Fix: hard-suspend the mic while speaking. The kind of problem you only find by building the thing, not speccing it.

I started with Claude Haiku driving the spoken responses, then switched to ElevenLabs because everyone said it was better. It is, and it was wrong for this. The voices were too dramatic, too "reading you a book," when a mission console needs clipped, neutral acknowledgment. I reverted. Lesson: "best" is genre-dependent; a control room and an audiobook are different products.

The biggest real performance win came later, retroactively, from AloInLaws: fixing the speech recognition language to English instead of auto-detect. I didn't know the option existed when I built SkyWatch. Auto-detect makes the system spend time guessing the language before parsing the command; a fixed language cuts that entirely. If I reopened this build, that's change number one.

Pinch-and-drag works decently in the conditions I built it in: one person, a clean background, decent light. Outside that, it degrades in predictable ways. Any second hand in the camera frame confuses the tracking. And in low light, the pinch, my equivalent of a mouse click, would get unreliable exactly when precision matters most. A real deployment would need either constrained mounting (a camera angled at a known operator position) or a confidence threshold that falls back to voice when tracking gets noisy. Knowing where an input modality fails is as much the product judgment as making it work.