Agents take over code generation. What grows in its place? The prose around code — context docs, instructions, architecture decisions. That’s a typing-bound task, but voice can unblock it.
Voice dictation isn’t for dictating code. It’s for the rich context that agents actually need: explaining intent, describing constraints, iterating on design. Speaking is about 3× faster than typing, and when input is cheap you naturally give the model more to work with.
A few patterns that work:
- Voice your long prompts, type your short ones. Surgical one-liners are fine to type. But when you need to explain context, constraints, and intent — just say it.
- Two humans discuss, AI listens. Have a design conversation with a colleague while AI synthesizes structure from the transcript.
- Pair voice with a screenshot. Drag in an image for grounding, speak for intent. High-leverage combo.
The catch: there’s an awkward-phase barrier. Five minutes isn’t a fair trial — it takes sustained use before the workflow clicks. And voice is for prompts, not for camelCaseIdentifiers.
Tools to try: macOS built-in dictation (Fn twice) works for a first test. For better accuracy, Handy is open-source, runs Whisper locally, and works across apps — choose whisper-large-v3 for Norwegian. (Wispr Flow is a polished commercial option but sends audio to some cloud.)
💡 Try this: next time you need to explain a non-trivial task to the agent, press the dictation key (Fn twice on Mac) instead of typing. See if the prompt comes out richer. 🔗 Peter Steinberger — “Just Talk To It”
How do you give input to your coding agent?
🟢 Voice regularly — for prompts and context
🟡 I’ve tried voice but dropped it
🔴 Typing only — I’m fast enough
⚪ Hadn’t considered using voice
What’s your biggest barrier to using voice for coding work?