Skip to content

// essays · 19 October 2026 · 3 min read

Typing is the bottleneck in agentic coding, not thinking

Claude Code already listens. Now it talks back, and you can interrupt it mid-sentence.

You talk faster than you type. Everyone does. A Stanford study found speech input was about three times faster than typing on a phone keyboard. That gap, not your reasoning speed, is the real bottleneck in an agentic coding session, where you spend the whole day describing what you want.

Claude Code already fixes one side of this. Turn on /voice, hold the spacebar, talk, and your speech becomes the prompt. What it doesn’t do is talk back.

So I built the other half.

It’s a small Claude Code plugin called voice-out. Every reply gets spoken aloud the moment it’s ready, using SAPI, the text-to-speech engine that comes built into Windows. No API key, no network call, nothing leaving the machine.

Code blocks get a short “code block” cue instead of being read out character by character. Tables, file paths and links get the same treatment. What’s left is just the prose, said out loud.

The dead air was the real problem, not the missing voice.

Here’s what an agentic session actually looks like most of the time: you ask something, then you watch a cursor while Claude reads files, runs tests, or searches the codebase. That’s dead time if you’re staring at the screen just to catch the moment it’s done.

Once replies speak themselves, you don’t have to. Step into the kitchen, glance at something else, keep moving. You’ll know exactly when it finished and what it found, because it told you.

Interrupting it was the actually hard part.

Claude Code has hooks: small scripts it runs automatically at set moments, like when a reply finishes or when you send a new prompt. It does not have a hook for “the user is holding a key down right now.” So when you want to cut a reply off mid-sentence, there’s no event to listen for.

The fix was a small background listener that’s always on. Hold the spacebar anywhere on the machine for 0.6 seconds (long enough that it’s clearly deliberate, not a typed space) and whatever is speaking stops immediately. The threshold is configurable.

A nice side effect: holding space is also how /voice starts recording. So the moment you start talking back, Claude stops talking. It feels like interrupting a person, which is the point.

That listener is deliberately global, not limited to Claude Code’s own window, because the alternative was no interrupt at all. The honest tradeoff: holding space in any other app past the threshold also counts. That’s a real cost. Not being able to shut Claude up mid-ramble was a bigger one.

One limit worth being upfront about.

Giving Claude Code a new instruction still needs a hand on the keyboard, because /voice records only while you hold a key. This plugin doesn’t touch that side. What changes is everything after you’ve asked: the answer comes back by ear, not by eye. That’s not fully hands-free. It’s half of it, done honestly, with the harder half left for later.

Why a plugin, not a skill.

A skill is a set of instructions Claude loads when it needs them. It can’t register a hook that fires every time a reply finishes. Only a plugin can ship hooks. voice-out bundles its own, so installing it wires up speaking (when a reply finishes) and stopping (when you send a new prompt) automatically. No editing your own settings file.

One setup command handles the rest: it installs the one Python package it needs, picks a voice already on your machine, and starts the background listener on login.

/plugin install voice-out --marketplace thisismairaj/voice-out
/voice-out:setup

That’s it.

p.s. It’s Windows only for now. The speech engine and the listener are both Windows-specific, and I’d rather say that plainly than let you find out the hard way.

Muhammad Mairaj is a Senior Backend Engineer with 8 years building payment and data systems. He writes about data engineering, measured. More about him.