I think better out loud. When something on a page is wrong, the quickest way to get the thought somewhere useful is to say it while pointing at it. Typing it out loses half of it. A screenshot on its own loses the other half.
So I set up two keys on my Mac. The Globe key is ordinary dictation: hold it, talk, let go, and the words appear wherever the cursor is. Shift and Globe together starts what I call a session. I talk about whatever is on screen, scroll, point with the mouse and switch tabs. When I let go, the words and a set of screenshots are saved together as one note. My assistant reads those notes, so when I say "move this down a bit" it can see what "this" was.
Everything in that loop runs on the Mac itself. The transcription model is local, the screenshots never leave the machine, and the notes are plain files in a folder. This article walks through setting it up. It takes about twenty minutes.
What stays on your machine, and what doesn't
"Entirely local" is the whole point, so here is exactly where it stops.
The recording, the transcription and the screenshots are all local. Superwhisper runs its speech model on the Mac. A small watcher script takes the frames and writes the notes. None of it touches the network.
The one thing that does leave is whatever your AI assistant reads. If you use Claude Code, the words of a note go to Claude, and so does any screenshot it chooses to open. It opens only the frames it needs, usually one or two, and those do go. If that matters for your work, a second part is coming. A small local model looks at the screenshots first, and only the ones it can't make sense of get passed on.
What you need
- A Mac on macOS 26 or later, with Apple silicon.
- Superwhisper, version 2.12 or later. That's the version that added a separate keyboard shortcut for each mode, which this depends on. The free tier is enough to try it.
- Apple's command line tools, for Python and a compiler. If
xcode-select -pprints a path, you have them.
Setting it up
1. Free up the Globe key
macOS uses the Globe key for the emoji picker and for its own dictation. In System Settings, Keyboard, set "Press Globe key to" to "Do Nothing", then open Dictation and turn it off. Tapping Globe should now do nothing at all.
2. Install Superwhisper
Pick a local model during setup so recordings stay on the machine. Record one short test, which creates the folder the watcher reads.
3. Make two modes
Superwhisper calls each way of recording a mode. You want two. One called Session, on Shift and Globe, with paste turned off. One called Dictate, on Globe alone, with paste turned on. Then clear the push-to-talk key under Shortcuts, and leave "Paste result text" off in the general settings.
The names matter. Only a mode called Session makes notes. Everything else stays plain dictation, so a quick reply in a chat box never turns into a note. Clearing push to talk matters too, because it records in whichever mode you used last. With a shortcut on each mode, each key always means the same thing.
4. Move your screenshots off the Desktop
A screenshot you take mid-session with Cmd+Shift+4 joins the note, but the watcher can't read the Desktop without full disk access. Move them instead:
mkdir -p ~/Pictures/Screenshots
defaults write com.apple.screencapture location ~/Pictures/Screenshots
killall SystemUIServer
5. Run the installer
Download it, read it, then run it:
curl -fsSL https://intelligenceintelligence.io/skills/voice-notes/install.sh -o install.sh
less install.sh
sh install.sh
It installs the watcher, builds a tiny app for the screen recording permission, writes a settings file and starts it at login. It's safe to run again to update.
6. Grant screen recording
In System Settings, Privacy and Security, Screen and System Audio Recording, switch on Voice Note, then restart the watcher:
launchctl kickstart -k gui/$(id -u)/dev.solarise.voice-note
7. Try both keys
Hold Globe in a text field and say a sentence. It should appear where you're typing. Then hold Shift and Globe, talk about the page in front of you, scroll a little, and let go. About ten seconds later there's a new folder in ~/Notes/voice/.
What a note looks like
Each session gets its own folder with the transcript in note.md and the pictures beside it. While you talk, the watcher takes a frame of the front window every second and a half and drops any that match the one before. The rest are listed in time order, with each sentence you said slotted in between. So the frames under "move this down a bit" show what was on screen as you said it, pointer included.
Superwhisper times each word it hears, so each sentence starts exactly when you started saying it. When a model doesn't give word timings, the watcher estimates them from word counts and pauses in the audio. That's usually within a second or two.
Frames are saved at half size, which is the size you see on screen. That's 120 to 250KB each, and up to about 2,000 tokens when an assistant opens one. Text stays perfectly readable.
Handing notes to Claude Code
This is the bit that makes it feel automatic. A small hook runs before every message you send. It hands the chat any new session notes, with the paths to their frames. I don't paste anything. I talk, let go, and type "latest note", or just carry on with what I was doing.
When a chat takes a note on, it marks it picked. Every other open chat then shows that note as one line. When the work is done it marks it handled, and the note records where it went. That stops two chats working on the same thought.
The setup for the hook, and everything the assistant needs to read notes well, is in the voice notes skill. Give it to your assistant and it can do most of this setup for you.
Things I learned
- Thinking out loud takes practice. My first notes were full of "let's also" and "I don't know, you figure that out". That's fine. The assistant reads for intent.
- Say "this" and point. The pointer is in the frame, so "this caption" is enough.
- Long sessions work. Some of mine run to several minutes and hundreds of frames, and the timeline means the assistant opens only the ones near each point.
Undoing it
launchctl bootout gui/$(id -u)/dev.solarise.voice-note
rm ~/Library/LaunchAgents/dev.solarise.voice-note.plist ~/.local/bin/voice-note ~/.config/voice-note.json
rm -rf ~/Applications/VoiceNote.app
Your notes stay in ~/Notes/voice/. Remove the hook from your Claude settings if you added it, and put the Globe key back in System Settings.
Coming next: a local model for the screenshots
The next part moves the last piece on to the machine. A small vision model, running locally, reads each note's frames first. It writes down what each point refers to: which part of the page, which project folder, what you seem to want done.
Small models manage "the user is pointing at the pricing table and wants it wider". Spotting a problem in a picture by themselves is harder, so it also says when it isn't sure. Only those frames go to a larger assistant. Everything else stays put.