Vibe Coding Is Just Dictation Wearing a Hoodie

I build a Mac dictation app. I have a vested interest in not lying to you about this.

What is "vibe coding," actually, if you strip the costume off it? It's voice going into a microphone, a transcription model turning that voice into characters, an editor receiving those characters, and an LLM doing something interesting with them. Three of those four steps were here in 1996. Dragon NaturallySpeaking did all of them on Windows 95. The new bit — the LLM in the middle — is genuinely new and genuinely impressive. The bit everyone is selling you, though, isn't the LLM. The LLM is an Anthropic API call. The bit being sold is the hoodie around it.

Andrej Karpathy was honest about this in the original tweet. He said his "vibe coding" workflow was Cursor Composer plus Sonnet plus "I just talk to Composer with SuperWhisper." Three products. The first one is the editor, the second is the model, the third — the unglamorous one that did the actual work of turning his mouth into a keyboard — is a dictation app. Every retrospective I read in the months after skipped past that third part. You can guess which part is the hoodie.

A short tour of the current costumes

Cursor sells itself as "the best coding agent." It is, in fact, a VS Code fork that calls a Claude API and accepts voice input. Strip the LLM and it's an editor. Strip the dictation and you're typing. The agent — the bit on the marketing page — is the part Anthropic actually builds. Cursor's contribution is the wrapper, the keybindings, and the venture round.

Cursor in May 2026. A VS Code fork plus an Anthropic API call plus a microphone. Screenshot via cursor.com.

Wispr Flow sells itself with the tagline "Don't type, just speak." That is, verbatim, the marketing slogan that Dragon NaturallySpeaking used in 1997. They have added a cleanup LLM pass on top of cloud transcription, a slick onboarding flow, and a $25M round at a $700M valuation. The actual product is dictation. They are selling dictation. They are not allowed to call it dictation because dictation is a $49 category and "voice-to-text AI" is a $12-a-month-subscription category.

Wispr Flow's hero in May 2026. "Don't type, just speak." Dragon NaturallySpeaking, 1997, called and would like a word. Screenshot via wisprflow.ai.

Cluely sells itself as "the #1 Undetectable AI for Meetings." It is meeting transcription, plus LLM summarization, plus a UI that hides itself from screen-shares. The hoodie here is the opacity — you're not allowed to know it's dictation, because if you knew, the entire pitch ("cheat on everything") would deflate into "take notes during your meetings." Which is what it does.

Cluely in May 2026. Meeting transcription plus a summary, but the costume is opacity. Screenshot via cluely.com.

The hoodie is what costs money

Notice the cost structure. The transcription bookend — the part that turns sound into characters — is free. OpenAI open-sourced Whisper in September 2022. It runs on any Apple Silicon Mac. It is better than any commercial transcription engine I have ever paid for, and I have paid for several. The model that does the actual heavy lifting of dictation costs zero dollars and runs on hardware you already own.

Every product in the tour above takes that free, local primitive and bolts something on top of it: a cleanup LLM (Wispr), an IDE (Cursor), an opaque overlay (Cluely). Each bolt-on has a real engineering cost — I'm not pretending they're scams. But the bolt-on is also where the subscription lives. You're not paying for the dictation. You're paying for the costume that lets the founders pitch it as something other than dictation.

What we built EmberType to be

I'm in the dictation business. I'm not pretending I'm not. There's no LLM in the middle of EmberType. There's no cloud transcription on the back end. There's no agent layer, no overlay, no "voice-to-text AI." It's Whisper, running on your Mac, wired into a hotkey, pushing text into whatever app your cursor is in. Cursor, Claude Code, Slack, Bear, Apple Mail, anywhere.

It is the cheapest version of the primitive that every product in this essay is wearing a costume around. $49 once, no subscription, no audit trail, no audio leaving your machine. You can pair it with Cursor and Claude Code and reproduce most of Karpathy's workflow without sending a single byte of your codebase narration to anyone's server. The voice layer is the part that survived the term going out of fashion — so own the voice layer instead of renting it.

The hoodie comes off eventually

The term "vibe coding" already got walked back. Karpathy demoted it himself at Sequoia AI Ascent 2026, swapping it for "agentic engineering" — which is also, when you squint, dictation plus an LLM plus a meeting with your security team. By 2027 it'll have a different name. By 2030 the costume budget will be larger than the entire dictation software market was in 2010.

None of which changes the primitive underneath. The mouth-to-machine pipeline is the part that's actually useful and the part that's actually been here the whole time. Build on the primitive. Skip the hoodie.