Just Talk to It
Over the last few months, I’ve stopped typing my prompts to LLMs and started dictating them instead. I open the agent, fire up Superwhisper, and just say what I want for a minute or two straight. The responses I get back are noticeably more accurate, and I spend a lot less time going back and forth with the model. I wanted to share why I think talking to these models works so much better than typing.
Giving the Model More Context
Talking lets me cover far more context than typing in the same amount of time. A Stanford and University of Washington study clocked English speech at 153 words per minute against 52 on a keyboard, about three times faster. However, the speed itself isn’t what matters most to me. What matters is that talking removes the friction of writing everything out. When I type, I tend to leave things out. For example, I’ll skip the backstory or the list of things I’ve already tried, because writing it all out is tedious. As a result, the model only gets a small portion of the context in my head. When I talk, it’s easy to include everything, and the model gives a much better answer when it has the full picture.
LLMs Perform Better With More Context
LLMs perform better when you give them relevant context. Prompting guides from the major labs all recommend giving the model the details it needs instead of assuming it can infer them. There are two well-studied techniques that are really just specific kinds of extra context, and both are far easier to do out loud.
Few-shot prompting means showing the model a couple of examples of what you want before asking. It works well, but typing out examples is a chore, so I usually skip them. When I’m talking, I’ll throw in an example without thinking about it, like “I want it formatted the way I said X earlier.”
Chain-of-thought means walking through the reasoning step by step instead of jumping to the answer, which sharply improves results on hard, multi-step problems. When I talk, I naturally narrate my reasoning. I’ll explain how I would approach the problem, which edge cases I’m worried about, and what I would check first.
Additionally, talking changes the style of my prompts. When I type, I tend to write in terse keyword fragments. When I talk, I explain things the way I would to a person. These models are trained on a vast corpus of human language, so natural, conversational phrasing is much closer to what they learned from.
Does More Context Always Help?
However, the goal isn’t simply to give the model more words. The extra context has to be relevant to the task. Research shows that stuffing a prompt with everything you can think of will degrade the output:
- Models tend to get “lost in the middle” of long inputs.
- Recall decays as the context grows.
- Irrelevant details can distract the model from the actual task.
Fortunately, the dictation apps help with this. They clean up the transcription by removing filler words and false starts, so the text that reaches the model is concise even when my speech wasn’t.
Getting Started
When Andrej Karpathy coined “vibe coding,” one detail was that he talks to his editor and “barely even touches the keyboard.” There are a few tools that make this easy to set up: Wispr Flow, Aqua, and Superwhisper, which I use mostly for the lifetime license instead of yet another subscription. It only took me about five minutes to get running.
If you work with LLMs regularly, I’d recommend giving dictation a try. Say the whole thought out loud instead of typing a compressed version of it. It’s been one of the most valuable changes to my workflow this year, and I hope it works as well for you as it does for me.
// END