Speaking Expands, Typing Compresses

A friend at a cafe told me to try voice dictation. He's a senior engineer, critical of tools by nature, so I took it seriously. That afternoon I downloaded Wispr Flow. Within a week I had the pro plan. It's become the most impactful tool I've used in the past year, outside of Claude Code.

I don't use it for Slack messages or emails or writing articles from scratch. For most communication, I still value typing, the deliberate word choice. But for one specific use case, voice has changed everything: prompting in AI conversations.

The insight here isn't about the software. It's about what happens when you remove the filter between thinking and input. When you type, you compress. When you speak, you expand. For AI collaboration, expanding is a cheat code.

The self-editing problem

When you type a prompt, you're doing two things at once: thinking and editing. You start a sentence, reconsider, delete half of it. You tighten as you go. You cut the tangent before it forms. By the time you hit enter, you've given the model the cleaned-up version of your thinking.

That's fine for polished writing. It's bad for prompting.

The problem: LLMs work better with more context, not less. The tangent you deleted might have been relevant. The half-formed thought you reconsidered might have clarified your intent. When you compress your thinking before it reaches the model, you're removing important human signals.

What voice changes

When you speak, you say things you wouldn't type because typing gives you time to reconsider.

The result is messier. More words, more tangents, more "actually, let me think about this differently." But that mess is your actual thinking. And AI can work with mess. It can find the thread. It can ask clarifying questions. What it can't do is recover the context you filtered out before you sent it.

More context, more nuance

I noticed this first in coding conversations. When I typed instructions for a feature, I'd give the essentials. When I spoke them, I'd naturally include why I wanted it, what I'd tried before, what I was worried about, how it connected to other parts of the project. The spoken version was three times longer and significantly more useful.

The nuance matters. "Add a button that does X" is different from "I've been thinking about adding a button that does X, because users keep asking about it in the feedback, but I'm not sure where it should go because the settings screen is already crowded, and I don't want to add another tab, so maybe it could live in the menu bar dropdown instead?" The second version gives the model everything it needs to make a good suggestion. The first makes it guess.

Unexpected directions

Speaking leads somewhere typing doesn't. When I talk through a problem, I often land in a different place than where I started. The act of articulating surfaces connections I hadn't made consciously. I'll start explaining what I want to build and realize halfway through that I'm describing something slightly different. Better, usually.

This doesn't happen when I type. Typing is too slow. I have time to stay on the original track. Speaking is fast enough to follow the thought where it wants to go.

For creative projects, this is invaluable. For coding projects, too. Some of my best architectural decisions came from voice-explaining a problem and noticing, mid-sentence, that I was solving the wrong thing.

The brain dump workflow

Voice dictation works especially well for longer inputs. I recently walked my dog for 30 minutes, talking through everything I knew about a topic. No structure, no editing, just a stream of thoughts captured by dictation. That brain dump became a prompt. I brought the transcript to an AI conversation, and we shaped it into the essay you're reading now.

The raw transcript was repetitive with half-finished thoughts. But it was also rich. Real experience, specific memories, genuine insights. Wispr Flow cleans up the filler words automatically, so what I got was my actual thinking without the "um"s and "you know"s. That cleanup matters. I wouldn't be as confident sharing raw speech otherwise. We went back and forth a few times, and an hour later, the essay was done.

I couldn't have produced this essay so fast by sitting down and typing a prompt. Not because I couldn't type the words. Because typing would have made me edit as I went, and I would have filtered out the rawness that made the context useful.

When to type, when to speak

This isn't about replacing typing entirely. Some prompts are better typed. Quick commands. Precise technical instructions. Anything where you know exactly what you want and just need to say it clearly.

But for anything exploratory in an AI conversation, anything where you're still figuring out what you think, anything where context and nuance matter: speak it. Give the model your actual thinking, not the compressed version.

The bottleneck moved

For a long time, I thought the bottleneck in AI collaboration was the AI. Get a better model, write better prompts, and learn the tricks. But the real bottleneck was upstream. It was how much of my actual thinking I could get into the prompt in the first place.

Voice dictation moved the bottleneck. Now the limit isn't input. It's having something worth saying. Which is a much better problem to have.

Parallel conversations

As I've leaned into this workflow, something unexpected emerged. I now work with multiple Claude Code terminals open simultaneously—one for each part of a project, or different projects entirely. I talk to each in turn. Give context to one, check on progress in another, think through a problem in a third.

The indie-hacker corners of the internet are full of people spinning up dozens of agents, orchestrating AI swarms, pushing the limits of parallel execution. I'm not doing anything that ambitious. But I've noticed my own behavior shifting. Where I once worked in a single conversation, I now naturally spread across several.

My wife hears me from the other room. She says it sounds like I'm on calls with colleagues, collaborating all morning. A strange sci-fi moment when you notice it. But the workflow feels human in a way I didn't expect. I'm not staring at a screen in silence. I'm talking through problems, explaining what I want, thinking out loud.

The productivity follows. An hour of speaking across multiple agents gets an immense amount done. Typing to three conversations at once would be exhausting. Speaking to them is just talking.

Voice dictation didn't just improve my prompts. It made parallel collaboration possible.

Late to the party

I'm not discovering something new here. Colleagues who use ChatGPT have been using voice mode for this exact reason. Engineers I know have been dictating into their IDEs for years. I was slow to try it. But now that I have, I'm fully on board for this specific use case.

While I use Wispr Flow, any good voice dictation works. The learning here is the expansion of speaking. It removes the self-editing filter. It lets you give AI more context, more nuance, more of your actual thinking. That's the unlock.

Try it for your next complex prompt. Don't type it. Say it out loud, capture it, send the whole thing. See what the model does with the expanded version of your thinking instead of the compressed one.

This is a working document. My workflow keeps evolving as I learn what's possible.