AI Creative OS
01One agent per editing style02Pick the agent by what you have03Write the look down in numbers04Voice first, with word timestamps05Build every scene as code, on the word06Make it check its own work07Ship it with the comment funnel armed08The 30-minute test

Four Editing Agents, One Prompt Each

Type a topic, get a finished reel. The four agents behind my reels (DOSSIER, VOX, ANCHOR and the Edit Agent): what each one makes, the pipeline they all share, and the prompt that builds your first one in Claude Code.
What you'll make
Want this running for you?Build it step by step inside AI Creative OS, with me in the thread.Get Your First Agent →
01THE IDEA
One agent per editing style
Each style is a folder Claude Code can run: STYLE.md (the look written down in numbers), AGENT.md (the stages, the exact commands, the checks) and template/ (the pipeline with every reel-specific value pulled into a plan file). You name the style and the topic, Claude Code reads the folder and runs it start to finish.
02THE FOUR
Pick the agent by what you have
DOSSIER: research-first explainers cut in thirds, your face, the real website, motion graphics. VOX: huge kinetic type with a marker on the key word, works with no face at all. ANCHOR: you small at the bottom, one motion-graphic card per sentence above you. Edit Agent: you hand it your own raw take and it cuts on the words, adds captions and proof. Topic only? Start with VOX. Recorded a take? Edit Agent.
03METHOD 1 · 20 MIN
Write the look down in numbers
Open a reel you love and step through it frame by frame: ffmpeg -i ref.mp4 -vf "fps=2,scale=240:-1,tile=6x3" sheet_%02d.jpg gives you contact sheets. Write ten numbers into STYLE.md: canvas, fonts, the accent colour, where the text sits, how a word enters (blur, rise, how many ms), how long a scene holds, which sounds play when. Numbers, not adjectives, or the agent guesses.
04METHOD 2 · 10 MIN
Voice first, with word timestamps
Write the script one line per on-screen beat, then voice it with the ElevenLabs with-timestamps endpoint so you get the start time of every word. That file is the clock for the whole reel: every card, highlight and sound is keyed to the word it belongs to, so the edit always lands on the voice.
05METHOD 3 · THE BUILD
Build every scene as code, on the word
A single HTML page with a render(t) function draws the frame at any time t; Puppeteer screenshots 30 frames a second and pipes them into FFmpeg. Let motion start about 0.12 s before its word and settle on a small spring, and give every beat one real visual (a screenshot, a logo, a clip, a chart), not just text.
06METHOD 4 · THE CHECK
Make it check its own work
Before anything is posted the agent pulls frames at fixed times into a contact sheet and looks at them, confirms the fonts loaded, that something visual is on screen in the first two seconds, that no text sits on the face, and measures loudness. Write those checks into AGENT.md as gates: a run that fails one is not done.
07METHOD 5 · SHIP IT
Ship it with the comment funnel armed
The last stage posts the reel and arms a comment-to-DM automation on the same post, so everyone who comments the keyword gets the asset in their DMs, plus one short follow-up the next day. I run this through Zernio. This page is the asset from exactly that funnel.
08DO THIS FIRST
The 30-minute test
Pick one reel you like, run Method 1 on it, then paste the prompt below into Claude Code in an empty folder. You should have a 15-second test render in about 30 minutes. If you do not, the style sheet is missing numbers: add them and run it again.
The prompt · paste into Claude Code in an empty folder
Build me an editing agent for ONE style, in this folder.
1. Read STYLE.md (my numbers for the look). Ask me for anything missing before you build.
2. Make AGENT.md with stages: script (one line per on-screen beat) → voice with ElevenLabs word timestamps → plan.json (every element keyed to a spoken word) → reel.html with render(t) → render frames with Puppeteer into FFmpeg at 1080x1920, 30 fps.
3. Motion: start each move 0.12 s before its word, settle on a small spring, one real visual per beat (screenshot, logo, clip or chart).
4. QC gates before you call it done: contact sheet at 0.3 / 1 / 2 s and one frame per beat, fonts loaded, nothing over the face, loudness -13 to -10 LUFS. Show me the sheet.
5. Run it once on this topic: [YOUR TOPIC], 15 seconds.
THE STACK
What the agents run on
Claude Code: reads the agent folder and runs every stage. ElevenLabs: your cloned voice with word timestamps. FFmpeg + Puppeteer (free): frames in, mp4 out, on your own computer. Zernio: posting to every platform and the comment-to-DM funnel.
Want this running for you?
Build it step by step inside AI Creative OS, with me in the thread.
Get Your First Agent →
Zdeno Mucina
Not ready for that?
Drop your email and 3 more guides open in the Library, free.