It's 11 PM. You're writing a design doc for a settlement bug fix, explaining the three systems involved and why they desynchronized at scale. You're speaking faster than you could type, because the mental model is sharp and if you switch to typing now you'll lose half the thread. And then the transcription stops. Hits its limit. And you're back to the keyboard, assembling fragments into prose, knowing the doc was clearer three minutes ago when you were still thinking out loud.
The Design Doc at Odd Hours
Design documentation happens when you have the mental model in your head. For backend engineers like Marcus at a Series B fintech in Stockholm, that's often after standup confusion or during a postmortem when the bug pattern is fresh and urgent. It's not a task you schedule. It's a moment when the architecture makes sense, and you need to capture it before the context evaporates. You reach for voice because your hands are on the keyboard, your brain is still running, and articulating the thought would move faster than typing.
The problem is structural. Cloud transcription services cap the free tier aggressively. Wispr charges $14 a month for uncapped access; Willow starts at $12; Superwhisper at $8.49. But the free tiers? 5,000 to 50,000 words per month. A single 5,000-word design doc in your first week of the month leaves you empty for the rest of it. Designers and PMs can work around this by batching shorter videos or using their phone sparingly. Developers can't. A design doc about payment settlement isn't optional. And when you hit the cap mid-sentence, the tool doesn't warn you. It just stops transcribing. You think you've got the full thought captured. You don't.
Why Voice Matters for Modern Development
The bottleneck in Cursor and Claude-era development isn't typing speed. It's explaining what you want the model to build. You're not dictating code. You're dictating architecture decisions, edge cases, performance constraints, trade-offs you've already considered. The more specific and complete your voice, the better the code suggestions. But that specificity is long-form. A voice note about settlement payment retry logic isn't 30 seconds. It's three minutes of context. A thorough explanation might be four or five minutes. On a keyboard, that's twelve minutes of typing.
When you hit a word cap mid-explanation, you lose the thread. You have to switch context: stop recording, edit what you captured, clean it up, then reconstruct the second half from memory. By then, the model's already seen your incomplete thought, and you're starting over with a new prompt. The mental stack you built is gone.
There's also a privacy layer. Marcus specifically switched to Cursor instead of VS Code because Cursor's tab-complete means fewer voice rewrites, but he absolutely refuses cloud-based transcription. Settlement logic contains customer data, rate configurations, system internals. He's not sending that to a transcription service, no matter how reputable. If voice writing requires cloud processing, it's off the table.
Free Plus Local Plus No Interruption
Recitey handles this differently. It runs Whisper locally on your device with zero per-word cost. There's no word counter. No metering. No paywall between your first voice note and your fiftieth. You speak, the model transcribes cleanly on your machine, the output lands in Slack or a design doc or a Linear issue description or a GitHub PR. No cloud roundtrip, no privacy audit to pass, no surprise bill when you exceed a free tier.
The difference is structural, not just economic. Whisper-large-v3 runs at high accuracy even on older hardware. A 5,000-word design doc transcribes in under 60 seconds on a mid-range laptop. You're not waiting for a cloud service to respond. You're not wondering if your traffic spike will trigger a rate limit. And your code stays on your machine.
This matters more than it sounds. Marcus refused cloud dictation entirely. But he's not an outlier. Any developer writing code related to payments, healthcare, finance, or security faces the same IP concern. The tools that ask you to send your voice to the cloud are, by default, off-limits for most serious development work.
The Trade-off
Local processing is slightly slower than cloud for polish. If you want a perfect paragraph, you'll still need to rewrite. Your first draft from voice is clean enough for code context, but not publication-ready. Recitey handles that in a second pass: takes the rough transcription and makes it read like prose. Turns "uh we need to uh handle the case where the payment times out after uh 30 seconds" into "If payment times out after 30 seconds, retry with exponential backoff." But that's the paid tier. The free tier is the foundation. The uncapped, unmetered foundation.
The question most tools ask developers is implicit: "How much voice are you willing to use if we charge you for it?" The better question is: "Why shouldn't voice be free to use as much as you need?"
The design doc doesn't wait. Neither should your transcription.