Build a workflow
Wire up a trigger, a few steps and an effect — then put it on a button.
Workflows live under Workflows in the companion. Start from one of the three built-ins rather than an empty canvas — Dictate to cursor, Voice Assistant, Meeting Recorder — and change what you need.
See Workflows for what the pieces are, and Bindings for how one gets fired.
Build it
Start from something that runs
Duplicate a built-in. Dictation is the shortest useful graph there is: capture audio, transcribe it, type it at the cursor.
Add steps
Steps connect only where their types agree — audio flows into a transcriber, text flows into the thing that types it. If a connection will not make, that is why.
Choose where the output goes
The last step is the one that changes your machine: type at the cursor, press keys, tell the device to do something, or send audio to the virtual microphone or a recording.
Bind it
Put it on a device button, a keyboard shortcut, or leave it manual. A workflow with no binding is inert — which is fine while you are still building it.
The bit that trips people up
Not every workflow reaches the device.
If the last thing your workflow does is press keys on your Mac, then it runs entirely on your Mac. The device sends the button press and takes no further part — so it will not appear as a binding on the device, and it will do nothing while the companion is closed.
If your workflow captures audio, the device is genuinely involved, because the device has to open the microphone. That one becomes a real binding on the device.
Adding a model
Drop in Ask agent and the text passes through a model on its way to wherever it was going. That is the whole difference between Dictation and Voice Assistant.
You need a model provider configured for that step to do anything — see Connect a model provider. Every other step works without one.