Should I vibe code
AI dictation for Mac that turns speech into text anywhere
Speech in, text out, nothing leaves the machine. Go.
?
Their verdict, the Pro price and the build-time estimate come from their entry, MIT-licensed. Checked 2026-08-03.
?
Our verdict, the regret score and everything below it. Editorial and unsponsored — nobody can pay to be moved.
The honest answer
why the verdict is what it is
The pipeline is a hotkey, a recorder and a local model. It is the kind of thing that takes an evening and then quietly works for years.
What actually breaks
not "if". the specific failures.
- Model size against latency, where the accurate model is slow enough that you stop dictating and the fast one is wrong enough that you stop trusting it
- Inserting text into whatever is focused, which means the app can type into any window including the wrong one
- Punctuation and formatting, which is most of what makes dictation usable and is not what transcription models optimise for
- Microphone permission and device switching, so plugging in headphones mid-session silently captures nothing
- Proper nouns and jargon, which are wrong in exactly the sentences you cared about
There is genuinely no incident here — nothing leaves the machine and nothing is at stake, which is why the verdict is an unqualified go. The honest small failure is a hotkey and a focused window: you trigger dictation, start speaking, and the text lands in the search field of the app you were actually looking at rather than the document behind it. Nothing is lost. You just watch a paragraph get typed into a filter box, one word at a time, with no way to stop it politely.
Is that you?
the verdict is a default, not a law
- The model runs locally and audio never leaves the machine
- You insert text only into a target you confirmed, or via the clipboard
- You have measured latency with the model you intend to actually use
- Audio is sent to a service you have not told the user about
- Text is inserted into whatever happens to be focused with no confirmation
- You have not tested what happens when the audio device changes mid-recording
- Recordings are kept without a retention decision
If you build it anyway
the checklist, then the prompt that enforces it
- Run the model locally. It is the entire premise — speech in, text out, nothing leaves — and it also removes every privacy question at once.
- Prefer the clipboard over synthetic keystrokes, or capture the target window at the moment you start and refuse to type into a different one.
- Measure end-to-end latency with the real model and real audio length. Choose model size against that number rather than against benchmark accuracy.
- Handle audio device changes explicitly — headphones plugged in mid-session is common and silently produces nothing.
- Post-process for punctuation and capitalisation, since raw transcription output is technically correct and painful to read.
- Delete audio after transcription by default, and say so.
Before you build a dictation tool, apply these and push back if I ask you to break them. 1. Run the speech model locally. Tell me this is the point of the tool — audio never leaves the machine — and that it also removes the entire privacy question rather than managing it. 2. Prefer putting the result on the clipboard over synthesising keystrokes. If I insist on direct insertion, capture the target window at the moment recording starts and refuse to type into anything else, so a focus change mid-sentence cannot dump a paragraph into a search box. 3. Measure end-to-end latency with the actual model and realistic audio length, and report the number. Choose model size against that, not against a benchmark score. 4. Handle audio input device changes during a session explicitly — detect the switch, and either follow it or stop with a clear message. Silently capturing nothing is the common outcome. 5. Post-process the transcript for punctuation, capitalisation and paragraph breaks. Raw output is accurate and unpleasant to read. 6. Let me supply a vocabulary of names and jargon to bias transcription, since those are the words most often wrong and most often important. 7. Delete recorded audio after transcription by default, and state the retention behaviour in the interface. 8. Provide an obvious way to cancel mid-dictation that does not insert anything. 9. Out of scope unless I ask: real-time streaming transcription, speaker diarisation, cloud sync, meeting recording.
That one keeps you out of trouble. For the prompt that actually builds it, canivibecodeit.com has one.
their build prompt ↗Or don’t build it
the boring option, and the way back out
$8.49 a month buys tuned models, good punctuation handling and reliable insertion across applications. This is one of the strongest build cases on the site though: local models are good now, the pipeline is short, and the hot take is right that there is nothing here to be careful about.
Nothing accumulates here — a dictation tool holds no state worth migrating, which is part of why it is such a comfortable build. Keep any custom vocabulary in a plain file and the rest is disposable.
Local transcription engine that makes a simple dictation clone realistic.
Questions
Why prefer the clipboard over typing the text directly?
Because insertion is the only part that can misfire visibly. Synthetic keystrokes go wherever focus currently is, which may not be where it was when you started speaking — and a paragraph typed into the wrong field is awkward to stop mid-flow. The clipboard puts you in control of the paste, at the cost of one keystroke.
Are local models actually good enough?
For dictation, yes — this is a genuinely solved problem locally now, and the trade is latency against model size rather than accuracy against a cloud service. The remaining weakness is the same everywhere: proper nouns and domain jargon, which is why supplying a vocabulary matters more than choosing a bigger model.
Every week, someone ships something they shouldn’t have.
New verdicts, the worst thing that landed in the trap, and the occasional incident report. No other email, ever.
Text-based editing is a beautiful idea sitting on top of very fiddly media plumbing.
Screen plus camera to a file is platform APIs. The polish is where the money went.
Hotkey, Whisper, paste at cursor. This is the vibe coding success story everyone quotes.
last reviewed 2026-08-03 · verdict is editorial and unsponsored · shared entry data from canivibecodeit under MIT · not legal advice