Voice and pointer review sessions for websites and videos, packaged for Claude.
> /dictate the onboarding pages Opening Dictate Studio… Point, press T, talk, press T. Then Finish → Send.
Dictate records what you say together with what you point at, then lines each sentence up with its element, frame, or timecode. Words like “this” and “here” resolve to wherever your cursor was when you said them.
On your site, the Dictate extension for Chrome pins whatever you point at and takes a screenshot. Claude gets the page, the element’s selector, its text and its HTML, so “this button” means one button.
T talk · AltA pin without talking
Click the picture to pin a point, drag to box an area, press I and O to mark a range. Every note carries its timecode and where on the frame you pointed, and the video pauses while you talk.
IO range · M marker · ←→ one frame
P1
B1
Whisper runs on your machine and turns your voice into word-timed text, even when one sentence switches languages. Before you send, you fix anything it misheard and void the notes you don’t want.
Press Send and Claude reads every note with its anchor and screenshots. It replies with a numbered plan of what it understood, then makes the edits, and stops to ask only when a note is genuinely ambiguous.
### 2. /dashboard/listings · 00:31 into the session Page: [Listings](http://localhost:3000/dashboard/listings) > Make this button bigger and green. - **Pin P2** → `button.btn-primary` "Add listing" - selector: `main .toolbar > button.btn-primary` - element box: 1012,148 · 164×40 px in a 1440×900 viewport - **Pointing** "this" → `button.btn-primary` "Add listing" - **Screenshot** `frames/0004-p2.jpg`
/dictate.You need Node 20 or newer and Chrome or Edge. Nothing to clone: the skill starts the newest Dictate from npm every time, so updates reach you on your next review.
Installs /dictate for Claude Code from skills.sh. npx skills update keeps it current.
npx skills add D7y1/dictate -gInstalls faster-whisper (with the CUDA libraries on an NVIDIA card) and downloads the speech model. It takes a few minutes the first time and is safe to run again. Skip it and Dictate uses the browser’s live captions instead.
npx @d7y/dictate setupOnly for reviewing websites, and Chrome only lets you do this step yourself. Setup and the studio open the page to add it, then confirm it’s in.
In any project. Say what you’re reviewing, or give it a video file. Claude opens the studio, waits, and starts working when you press Send.
/dictate the onboarding pages
/dictate C:\footage\ep12.mp4Every action can be rebound, including to the middle button and both mouse side buttons, so you can review without touching the keyboard.
The server, the studio and Whisper run on your machine. The studio is plain files, so it starts instantly and has nothing to break between projects.
Tracks the page, pins elements, takes screenshots. Never records what you type into fields.
Zero dependencies. Serves the studio, collects every event, and writes the session folder.
ES modules with no build step: the rundown, the program monitor, the review screen.
Word-timed text on your GPU, warmed up while you record.
Opens the studio, waits for Send, then reads report.md and the frames.
Each session lands in your project under .dictate/sessions/ with report.md, session.json, the frames, the transcript and your audio. The folder ignores itself in git.
Open a project in Claude Code, type /dictate, and point.