When you switched to using LLMs for code, your keyboard workload didn't go down. It shifted. You're still typing constantly, but now it's not code, it's prompts, design docs, PR descriptions that explain what the model should build. The typing speed that used to matter stopped mattering. The real friction is drafting intent accurately without losing the thought mid-sentence. That's where most voice tools sell you short.
The shift from code to intent
Developers used to optimize for typing code. Keyboards, ergonomics, IDE shortcuts, all built around speed at the terminal. But LLM-first workflows inverted this. A backend engineer at a Series B fintech now spends half the day in Cursor or Linear, dictating design decisions, incident postmortems, and PR context to Claude or Copilot. The bottleneck moved from "how fast can I write code" to "how fast can I articulate what I want the model to build."
Voice makes sense here. You think faster than you type. Speaking the architecture of a payment settlement bug fix takes three minutes; typing it takes fifteen. But every commercial voice tool kills your flow with the same limitation: a word cap on the free tier.
Why word caps break the workflow
Picture this: You're finishing a detailed design doc at 11pm. Sixty minutes in, voice dictating a complex settlement architecture into Notion. The system counts silently. At 2,000 words, or 5,000, depending on the tool, dictation stops recording. Your voice keeps flowing. Your thoughts don't. You're either typing the rest out by hand or saving it for tomorrow or upgrading to paid.
This isn't a speed problem you can solve by talking faster. This is a workflow interruption that costs coherence and thinking flow.
Wispr Flow, Willow, Superwhisper, the entire category of voice-first tools monetizes the free tier by metering words per day. The assumption is that voice is a premium feature and the meter justifies the cost. But that assumption breaks when voice isn't about convenience, it's about thinking at full speed.
What local-first actually means
Most developers don't trust cloud dictation. Your code snippets, your architecture decisions, your internal postmortems, all shipped to a server you don't control. Some tools claim encryption, but the model still runs remotely. Data still leaves the device.
Local-first changes the equation. Whisper running on your machine means your voice never touches a cloud endpoint. No API call, no transcript sent anywhere. The only thing that moves is the text you approve. And because it runs locally with zero variable cost per word, there's no structural reason to cap the free tier.
Whisper-large-v3 achieves 95%+ word accuracy across languages on consumer hardware. That's the base layer. No metering needed.
From speech to polished prose
Local speech-to-text handles volume. What it doesn't do well is prose. Raw Whisper output is conversational, umms, false starts, run-on sentences. Fine for notes, not for a PR description or design doc your team reads.
That's where the cloud rewrite layer kicks in. Recitey's Pro plan includes a polishing pass that takes raw dictation and shapes it into structured prose, headlines, bullet lists, proper grammar, in under two seconds. The philosophy is: speech handles volume, the cloud handles quality. Free tier gives you unlimited dictation. Pro adds the rewrite you'd otherwise do by hand.
This separation matters. You're not paying per word for the thing you're actually doing, speaking. You're paying for the optional refinement that comes after.
The indie builder and the skeptic
Marcus, a backend engineer at a fintech, refuses cloud-based transcription on principle. His company's policies around code IP make him uncomfortable uploading snippets to third-party services, even with NDAs. He switched to Cursor specifically because its tab-complete reduces the need for voice rewrites, fewer wrong assumptions mean fewer spoken corrections.
When he discovered a local-Whisper solution with no word cap, it solved a problem he didn't know he had: the permission structure finally matched the workflow. He could dictate incident postmortems, design decisions, and Slack explanations without negotiating a metered cap. The tool got out of the way.
This is what "local-first" actually means to the developers who care about it. Not speed. Permission.
The bottleneck in LLM workflows is clarity of intent, not typing speed. Tools that still charge by the word or cap the free tier misunderstand the problem they're solving.