I keep pasting screenshots as text descriptions for my agent, which probably loses half the signal. Maybe the real move is picking a vision-capable model for UI work and a text model for pure code edits, instead of defaulting to one for everything. Routing by task type feels like the next thing worth setting up.