AI Creative OS
01What is in the zip02Install once03Drop in a raw take04It transcribes and cuts on the words05It drops in the numbers and draws the map06It checks, renders and fixes the audio07The 30-minute test

The Editing Agent

One raw take in, one finished reel out, and you never open a timeline. The Claude Code blueprint that made the Raw vs Edit reel: it transcribes, cuts on the words, drops in the numbers, draws the map, adds the captions, checks itself and renders.
What you'll make
Build it step by stepInteractive · ~3 min · with Z’s build guideStart →
0100 · THE FILES
What is in the zip
Five folders. editing-agent is the workflow: a SKILL.md Claude Code follows step by step, a captions script, the audio post script with the cold-open sound, and template/, the exact HyperFrames project the Raw vs Edit reel was rendered from. The other four are the editing styles as skills: raw-vs-edit, glass-hud, behind-head, kinetic-doc, each with its design spec, the working overlay component and a showcase. Unzip everything into ~/.claude/skills/.
0201 · SETUP · 10 MIN
Install once
You need Node 22+, FFmpeg and the HyperFrames CLI (npx hyperframes doctor tells you what is missing). Unzip the folders into ~/.claude/skills/, restart Claude Code so it loads them, then copy the template out as your first project: cp -r ~/.claude/skills/editing-agent/template my-reel.
0302 · INGEST
Drop in a raw take
Record one talking-head clip, 15 to 30 seconds, phone is fine. Put it at assets/take.mp4. Portrait or landscape both work: for the stacked 16:9 cards a portrait take gets a blurred fill behind a face-and-shoulders window, the same treatment you see in the reel.
0403 · THE CUT
It transcribes and cuts on the words
The agent runs npx hyperframes transcribe (local Whisper, no key) to get word timestamps, groups them into 2 to 4 word captions, then reads the transcript and marks one proof element per spoken beat: a punch-in on the verb, a counter on the number, the platform map on the 'everywhere' line, a metric card on the claim, the keyword pill at the end. Never more than one element per cue.
0504 · THE PROOF
It drops in the numbers and draws the map
The counter card counts up to a real number from your own accounts (reels shipped, automations armed, platforms), never money. The node graph draws lines from your logo to every platform the post goes to. The bottom card stays raw the whole time, so every overlay in the top card reads as evidence.
0605 · THE OUTPUT
It checks, renders and fixes the audio
npx hyperframes check until zero errors, snapshots at every beat so you can look, then a high-quality render and the post script: compressor, loudness to -13 LUFS, the intro sound under the first 2.6 seconds. You get the mp4, a contact sheet and a cover frame with eyes open.
0706 · DO THIS FIRST
The 30-minute test
One 20-second clip, the raw-vs-edit style, the prompt below. If it takes longer than 30 minutes, your clip is too long or HyperFrames is not set up; run npx hyperframes doctor again. Once it works, every talking-head clip you record from now on comes back edited.
The prompt · paste into Claude Code inside your copied template project
Run the editing agent on assets/take.mp4 in the raw-vs-edit style.
1. Transcribe with word timestamps and build 2-4 word captions with scripts/captions.mjs.
2. Put one proof element per spoken beat: punch-in on the verb, counter on the number (use a real count from my accounts, never money), node graph on the 'everywhere' line, metric card on the claim, COMMENT RAW pill at the end.
3. Keep the bottom card raw. Give the composition a 0.9s pre-roll with first-frame posters under the videos.
4. Run npx hyperframes check until 0 errors, show me snapshots at every beat, then render high quality and run scripts/post.sh.
Want this running for you?
Build it step by step inside AI Creative OS, with me in the thread.
Implement this today →
Zdeno Mucina
Not ready for that?
Drop your email and 3 more guides open in the Library, free.
Download the Editing Agent (zip) ↓ raw-vs-edit.zip (style only)