Notes on working with AI
Dictate instead of typing, free and private
Speaking is much faster than typing, and spoken prompts carry more context. Handy is a free app that turns your voice into text, offline.
Sixty seconds of talking, one page of text
You press a key, talk for a minute, let go, and the text lands in whatever app you were using. Email, Word, WhatsApp on your desktop, the Claude chat box. Nothing was typed, and the audio never left your computer.
That is Handy, a free, open-source dictation app for Windows, macOS and Linux. It is MIT-licensed and has around 37,000 stars on GitHub.
How it works
Handy runs a speech-to-text model on your own machine. You pick the model once, when you set it up, and that download is the only time it needs the internet. After that you press a shortcut (push-to-talk by default, and you can change the keys), speak, release, and the transcript is pasted wherever your cursor is. It does not need to connect to your other software, because pasting works everywhere.
There are two families of models. Whisper comes in Small, Medium, Turbo and Large, is multilingual, and handles Hebrew. Parakeet V3 is fast even on an ordinary processor and detects the language by itself, but for Hebrew you should pick a Whisper model. Bigger models are usually more accurate and slower, so start with Medium and move up if you are fixing too much.
The surprise is the price tag and the privacy. Paid dictation services typically send your voice to a server and charge a monthly fee. Here the recording stays on your machine and the cost is zero. If you dictate figures or talk about employees, that is the part that makes the conversation with IT short.
The real gain is not speed
Most people speak much faster than they type. But what I care about is what happens to your prompts.
When you type a request to Claude, you compress it. Every sentence costs effort, so you cut the background. When you speak, you explain. You mention the context, the exception, the person who will read the result. Claude works from what is in the conversation, so a richer spoken prompt gets a better answer from the same model, with no tricks.
Three ways to use it
A long brief for Claude. Talk loosely about this month's close, then end with:
Organize what I just said into a brief: which numbers moved, the likely reasons, and which questions are still open.
Two minutes of rambling comes back as a structured page, the kind you would never have typed out.
Email replies. In HR, you speak the answer to a candidate who didn't get the job, including the feedback you want to give, and ask Claude to make it kind and short. Feedback is the part that gets cut when typing, because it takes the most words.
Notes right after a meeting. In a small business, the minutes after a supplier call are when the details are sharpest. Speak what was agreed, who does what and by when, and Claude turns it into a task list with owners and dates. It's a note you dictate, not a recording of the call.
What to check
Transcription is good, not perfect. Names and numbers are where it slips, so read the text before you send it. If Hebrew comes out rough, try a larger Whisper model, and make sure your microphone is close. Larger models need a reasonably capable computer.
What to do this week
Download Handy from github.com/cjpais/Handy/releases (the project site is handy.computer). Pick a Whisper model, set your shortcut, and for one day speak your Claude requests instead of typing them. Compare how much more you said.