Clip Scribe

Live

Transcribe fake meetings and dress fake clips with masks and bubbles.

Not published

It passed its checks, but publishing was off where it was built, such as a local run, so it has no live link. Its recorded demo is in the panel.

Each line is what the check ran on, and no more.

This is a model's opinion, not proof.

Nothing here should be taken as tested.

Not checked Whether its answers are true
Not checked ChatGPT: only a person with an account can test it
Not checked Installing the Claude plugin or the Muse gadget pack on a real device
Not checked No Prove it run: nothing it comes with was run for real

No reference files were given. The brief asks for transcribing a made up meeting and annotating a made up clip with masks, stickers and captions.

Voice TranscribeTurns the meeting recording into text with turn timing in transcribe_meeting
Muse WonderCleans the transcript, assigns speaker labels and pulls out keywords in transcribe_meeting
SAM 3.1Finds and tracks objects in the clip from short noun phrases in mask_clip_objects
Muse ImageGenerates the transparent sticker images in make_clip_pack

Speaker labels are anonymous A and B from diarization and can be wrong or missing. The user names them, and a plain transcript is the fallback.
Recorded files only, up to 10 minutes, not live streams. The browser converts every file to mono 16-bit WAV first.
SAM takes one short noun phrase per call, no lists or negation, at most 16 objects.
Clips must be MP4 or MOV and stay short (5 to 20 seconds). Masks can miss fast or hidden objects.
Stickers are generated images sequenced in the browser, not true video generation.
Speech bubbles are drawn in the browser from transcript timing and SAM tracks.
Preview exports as WebM, recorded in real time. It is not an H.264 MP4.

Add it in the Muse app as a custom connector. It exposes 3 actions: transcribe_meeting, mask_clip_objects, make_clip_pack.

Muse Code made 6 tool calls and 2 edits across 2 files (+7 −0).
npm run verify passed on the first try.
Screenshot judge: pass. Final Canvas shows dressed clip with speech bubble, 2 tracked persons and 2 transparent stickers.
Built in 11 min for about $0.01 in model calls (est.).
Publishing is off for this run (it's a Vercel preview, and a product published here would deploy into production's Vercel projects), so the product didn't get its own link. The preview, checks and demo above ran for real.
How it was built: 193 tool calls, 58 edits touching 58 files { <read-skill-result name="muse-app-ui" status="ok"> <read-skill-result name="meta-voice" status="ok"> <read-skill-result name="meta-reasoning" status="ok"> <read-skill-result name="meta-sam" status="ok"> <read-skill-result name="meta-image" status="ok"> <read-skill-result name="recipe-voice" status="ok"> <read-skill-result name="recipe-sam-media" status="ok"> <read-skill-result name="verify" status="ok"> <read-skill-result name="product-ui" status="ok"> { {"items":5,"ok":true,"revision":1} { Edited `/tmp/standup.txt` { { (failed) { Edited `/tmp/genclip.mjs` {