How I Use Voice Dictation to Control My AI Agent (Typeless + OpenClaw)

I talk to my AI agent more than I type to it. That probably sounds weird, but hear me out — once you try voice dictation for agent workflows, going back to typing feels like writing letters by hand in the age of email.

I've been running OpenClaw as my personal AI agent for a few months now. It handles everything from file management to web research to automating my dev workflows. The thing is, working with an AI agent means writing a lot of natural language — prompts, instructions, configuration files, memory notes. And typing all of that out? Painfully slow.

That changed when I started using Typeless for voice dictation. My interaction speed roughly tripled. Not exaggerating — I timed it.

The Problem: Too Much Typing for "Natural Language" Interfaces

Here's the irony of AI agents: they're designed to understand natural language, but we still communicate with them by typing. Think about that for a second. You have this powerful agent that can parse complex instructions, and you're pecking away at a keyboard at 60-80 WPM when you could be speaking at 150+ WPM.

With OpenClaw specifically, there are several places where you write a lot of text:

• SOUL.md — the personality and behavior configuration for your agent. This can easily be 500+ words of detailed instructions about tone, preferences, and rules.
• Prompts and instructions — when you ask your agent to do something complex, you want to be specific. "Refactor this function" is fine, but "Refactor this function to use dependency injection, add error handling for the three edge cases we discussed yesterday, and make sure the return type matches the interface we defined in types.ts" is better.
• Memory files — daily notes, context dumps, things you want the agent to remember across sessions.
• Chat messages — especially when you're going back and forth debugging something or brainstorming.

I was spending 30-40% of my "agent time" just typing instructions. That's not productive.

Enter Typeless: Voice-to-Text That Actually Works

I'd tried voice dictation before — Siri, Google's built-in stuff, various Chrome extensions. They all had the same problems: mediocre accuracy, no punctuation handling, and they'd choke on technical terms.

Typeless is different. It uses AI-powered speech recognition that actually understands context. When I say "open bracket function name colon string close bracket," it knows I mean (name: string). When I say "new line," it adds a line break. When I pause, it adds appropriate punctuation.

The accuracy is genuinely impressive. I'd estimate 95%+ for regular English and around 90% for mixed technical/conversational speech. That's good enough that I rarely need to go back and fix things.

My Actual Workflow: Voice + OpenClaw

Let me walk through how I actually use this day-to-day.

Configuring SOUL.md

My agent's SOUL.md is about 800 words of personality configuration, behavioral rules, and preferences. Writing that by typing took me about 25 minutes the first time. When I rewrote it using Typeless, it took 8 minutes. I just... talked through what I wanted my agent to be like, as if I was describing it to a friend.

"I want the agent to be direct and concise. No fluff, no corporate speak. When it doesn't know something, it should say so instead of guessing. For code-related tasks, always show the diff before applying changes. For research tasks, cite sources with links..."

It flows naturally because you're literally just describing what you want. That's the whole point of natural language configuration — and voice makes it feel like actual natural language instead of carefully typed prose.

Writing Complex Prompts

This is where voice dictation saves the most time. When I'm asking my agent to do something multi-step, I used to spend time carefully composing the prompt. Now I just talk through it:

"I need you to look at the API response handler in the payments module. The error handling is inconsistent — some endpoints return error objects, others throw exceptions. I want you to standardize everything to use the Result pattern we set up last week. Check the three main endpoints: create payment, refund payment, and get payment status. For each one, wrap the external API call in a try-catch and return a Result type with either the success data or a typed error."

That took me about 20 seconds to say. Typing it would have taken over a minute, and I probably would have been less detailed because typing long instructions feels tedious.

Mobile Conversations via Telegram

This is the game-changer. I have OpenClaw connected to Telegram, so I can chat with my agent from my phone. Typing complex instructions on a phone keyboard? Miserable. But with Typeless running on my phone, I can dictate detailed instructions while walking my dog, commuting, or lying on the couch.

I've literally configured deployment pipelines from my phone while making coffee. Not because I'm a workaholic — because it was easy enough that I could do it in the 3 minutes while the water was boiling.

Writing Memory Files

At the end of each work session, I dictate a quick summary into my agent's memory file. "Today I worked on the authentication refactor. Got the JWT validation working but still need to handle token refresh. The edge case with expired tokens during active sessions needs a queue-based approach — talked to the team about this and we agreed on the retry pattern from RFC 6585."

Thirty seconds of talking vs. two minutes of typing. And because I'm speaking naturally, I tend to capture more context and nuance than when I type, where I unconsciously abbreviate everything.

The Numbers

I tracked my usage for two weeks to get real data:

• Average typing speed for prompts/instructions: ~70 WPM
• Average speaking speed with Typeless: ~160 WPM (accounting for corrections)
• Net speed improvement: roughly 2.3x for short messages, 3x+ for longer instructions
• Time saved per day: approximately 25-35 minutes on days with heavy agent interaction
• Accuracy: high enough that I correct maybe 1 in 20 sentences

The speed difference is more dramatic for longer inputs. A quick "check git status" doesn't benefit much from voice. But a 200-word detailed instruction? Voice wins by a landslide.

Tips If You Try This

A few things I learned the hard way:

1. Speak in complete thoughts. Typeless handles pauses well, but if you stop mid-sentence to think, you might get weird punctuation. Better to pause, think, then speak the full sentence.

2. Use it for first drafts. Voice dictation is amazing for getting ideas out fast. I'll dictate a rough version of a SOUL.md section, then do a quick editing pass with the keyboard. Faster than typing from scratch.

3. Train yourself on technical terms. Typeless picks up on context, but for very specific terms (library names, custom types), it helps to speak clearly and slightly slower.

4. Don't whisper. I tried using it in a quiet office once by whispering. Accuracy dropped significantly. Normal conversational volume works best.

Why This Matters Beyond Convenience

There's a deeper point here. AI agents are supposed to make us more productive by understanding what we mean. But if the input bottleneck is typing speed, we're leaving a lot of that potential on the table.

Voice dictation isn't just faster — it changes how you interact with your agent. You give more context, more detail, more nuance. You describe problems the way you'd explain them to a colleague, not in the compressed shorthand of typed messages.

The combination of Typeless for input and OpenClaw for execution has genuinely changed my daily workflow. I spend less time typing and more time thinking about what I actually want to accomplish.

And honestly? Talking to your AI agent feels more natural than typing to it. Because it is.

---

If you're into AI tools and productivity workflows, I write about this stuff weekly in my newsletter: AI Product Weekly. No spam, just tools and techniques that actually work.