Until recently, typing was the default way to write. As voice-to-text has improved, the keyboard’s limits are easier to see.
Typing makes you edit while you think. You pause, rephrase, and reorganize in real time. That can improve clarity, but it can also cut incomplete or exploratory ideas before they reach the page.
Dictation works differently. You can speak more freely, change direction mid-sentence, and capture ideas as they form. Most people also speak faster than they type.
For years, dictated text was too messy for serious writing. Reliable AI cleanup changed that.
Dictation As Raw (But Complete) Intent
When you prompt an LLM for coding tasks, output quality depends on context and detail.
Typing friction often keeps us from including the details the model needs.
With voice, adding detail is easier. You naturally explain why you want a specific implementation, mention edge cases, and describe what you already tried.
Spoken prompts do not need to be clean. They need to be complete.
Raw dictated text is usually unstructured. A meta-prompt acts like a compiler for intent, restoring structure.
Instead of treating speech as the final prompt, treat it as source material and transform it into a structured prompt with:
- Task structure and output format
- Exact file mapping from vague references
- Repository conventions and constraints
At Utter, dictated text can run through custom AI prompts before insertion, adding structure and project context. Learn more in Utter AI prompts.
Why This Works
The benefit is not just speed. It is signal density.
Voice captures caveats, uncertainty, rationale, and implicit constraints people often skip when typing.
For coding workflows, that can mean:
- Better scoped tasks
- More accurate model behavior
- Fewer missed assumptions
Tradeoffs
Voice is high-bandwidth for intent but weaker for precision work. You still want a keyboard for exact edits like regex tweaks, identifier changes, or line-level refactors.
Environment also matters. Noisy spaces reduce dictation quality.
Meta-prompting adds complexity too. As prompt systems grow, debugging gets harder.
Still, the pattern is clear:
Lower friction changes how much people say. That changes outcomes.