Dictate

Point, talk, submit. Dictate records you giving notes on a video or a website (your voice, your cursor, what you click, what you pin) and hands Claude a session it can act on without asking "which one?".

Adding a note is one key: click (or just hover) the thing, press T, talk, press T again. The mic is already open, so it starts the instant you press. Hold T instead if you prefer push-to-talk. Esc throws an open note away.

Or use the mouse only: the mouse's back side button does the same as T (tap to open and close a note, or hold it), and the forward side button plays and pauses the video. Every control can be rebound on the home screen under Controls: click Add, then press any key, the middle button, or a side button. While a session runs, Dictate stops those buttons from navigating back or forward, in the studio and on your site.

• Video review: every note lands on the timecode with a pin where you clicked. Drag to box an area, I/O to mark a range to cut. The video pauses while you talk.
• Web review: browse your platform page to page. T works inside your site's pages. The Dictate extension pins the element under your cursor, records the page, what you click and select, and takes screenshots. While a note is open, clicks pin elements instead of following links.
• Show what you like: in the middle of either review, open any other site (Dribbble, a competitor, a YouTube edit) and talk there, or click Screenshot and drag around the part you mean. Dictate tags that site Reference, so Claude treats it as inspiration rather than your code, applies it to the page you were on (or the timecode you paused on), and gets the exact font, colors, spacing and radius of what you cropped.
• Submit: Whisper turns your voice into word-timed text on your GPU (Arabic and English, even mixed). You fix anything it misheard, void notes you don't want, then Send. Claude gets a report with every note anchored to a timecode or a DOM element, plus the images.

Install

npx skills add D7y1/dictate

That's the install. Type /dictate in Claude Code and start talking. The installer asks whether to add it to this project or all of them; -g picks all of them. You need Node 20+ and Chrome (or Edge).

What happens on your first review

• Speech to text sets itself up in the background. Dictate puts faster-whisper in its own Python environment (~/.dictate/python, so your Python isn't touched), adds the CUDA libraries on an NVIDIA card, and downloads the speech model. That's up to 3 GB, once. Live captions cover until it's ready. It needs Python 3.9+; without it Dictate stays on live captions and the studio tells you. Set DICTATE_NO_WHISPER=1 to turn this off.
• The extension, if you review a website. Chrome only lets a person add one, so the studio walks you through it, once. Video review doesn't need it.

Prefer to do it all up front, or want to retry something? npx @d7y/dictate setup runs the same steps in your terminal. It's safe to re-run; it skips anything already done and tells you what's left. Flags: --no-whisper, --no-model, --no-open.

Updates

Nothing to do. /dictate starts Dictate with npx @d7y/dictate@latest, so every review runs the newest release.

• The skill: npx skills update.
• The extension: from the Chrome Web Store, Chrome updates it by itself. Loaded unpacked, Dictate refreshes ~/.dictate/extension when a release changes it; press the reload arrow on its card in chrome://extensions.

Remove

npx skills remove dictate, remove the extension in chrome://extensions, and delete ~/.dictate (that includes Whisper and its Python environment). The speech model stays in ~/.cache/huggingface until you delete it.

Develop from source

git clone https://github.com/D7y1/dictate.git
cd dictate
npm run setup

Or double-click setup.cmd. From a clone, setup points the /dictate skill at your folder (so your edits are what /dictate runs), adds a dictate command with npm link, and has you load the extension unpacked from extension/. Run npx @d7y/dictate setup to switch back to the published version.

Use it

From Claude Code (any project): type /dictate, optionally with what you're reviewing ("/dictate the onboarding pages", "/dictate C:ootageep12.mp4"). Claude opens the studio, waits, and starts working when you hit Submit: first a numbered plan of what it understood, then the edits.

By hand:

npx @d7y/dictate                                   # studio home; sessions go to ./.dictate/sessions
npx @d7y/dictate --video D:cutsep12.mp4
npx @d7y/dictate --url http://localhost:3000 --out C:projectssite.dictatesessions
npx @d7y/dictate --latest                          # print the newest session folder

(With npm install -g @d7y/dictate, or from a clone after setup, it's just dictate.)

Then tell Claude "read my latest dictate session".

Video review keys

These are the defaults; everything except frame stepping (arrows, J, L, Backspace) can be changed under Controls.

This is the beginning of the README. Read the rest at the original. Read the full README ↗