How it works
An AI video editor is just Claude with three things: the right tools, the right libraries, and a clear picture of your style. Most setups only give it the tools. That's why the edits look like AI slop.
- A folder it works in. Raw footage goes in one side, finished videos come out the other.
- 4 free tools. Remotion builds the video, FFmpeg cuts it, Whisper writes the captions, yt-dlp grabs videos to study.
- 2 libraries. Real sound effects, and 200+ motion graphics with the prompts behind them.
- Your style. It studies 5 to 10 of your videos and 2 or 3 from creators you like, and writes your style guide.
- The edit. You drop in a raw take and your script. It cuts, captions, adds graphics, sound and music, and shows you a preview before it exports anything.
You don't install any of it yourself. One prompt does it all.
What you need
- Claude Code The Code tab in the Claude desktop app. Any paid plan.
- A Mac Windows works too, the prompt adjusts. A Mac is the smoothest.
- About 5 GB of free space Tools, Whisper's model and your footage.
- 30 minutes Most of it is downloads. You just say yes.
- Your videos 5 to 10 of your best ones, plus 2 or 3 from creators you like.
1. The build prompt
- Download the Claude desktop app and open the Code tab.
- Make an empty folder called AI Video Editor in your home folder (not on an iCloud Desktop).
- Start a new chat, pick Local, and select that folder. Now Claude can see and use the files in it.
- Paste the prompt below. Say yes when it asks to install things. When it prints READY, you're set.
You are going to turn this folder into my personal AI video editor for short-form videos (Reels, TikTok, Shorts, 1080x1920). I am not technical. Do every step yourself, explain what you're doing in one plain sentence per step, and only stop to ask me when you truly need me (a password prompt, a yes/no, or my videos). STEP 1: FOLDERS First check where this folder lives. If it's inside iCloud Drive (Desktop or Documents with iCloud sync on), warn me: iCloud offloads files to save space and that breaks renders. Suggest moving it to my home folder (~/AI Video Editor) and wait for my answer. Create this structure inside the current folder: - raw-footage/ (I drop my raw takes here) - my-videos/ (5-10 of my own best videos, so you learn my style) - inspo/ (2-3 videos from other creators I want to borrow ideas from) - library/sounds/ (real sound effects) - library/music/ (music I'm allowed to use) - library/motion/ (motion-graphic references + the prompts behind them) - style/ (my style guide lives here) - jobs/ (one folder per video you edit) - exports/ (finished videos) Then create a CLAUDE.md in the root that explains this folder to any future Claude session: what each folder is for, the workflow in STEP 6, and the house rules in STEP 7. STEP 2: TOOLS Check what's already installed and only install what's missing. I'm on a Mac (if I'm not, use the right installer for my system). Install Homebrew first if it's missing. - ffmpeg: cuts, trims, joins and exports video and audio. On a Mac install `ffmpeg-full` from Homebrew (brew install ffmpeg-full) and put it first in my PATH, because the plain version can't handle iPhone HDR footage (it's missing the zscale filter). Check zscale is there. - Whisper (OpenAI's open-source version, running locally on my computer, free): turns my speech into word-by-word timestamps for captions and for finding dead air. Download the "base" model, and "medium" if my computer can handle it. - yt-dlp: downloads videos from a link so you can study them. If Instagram blocks it, tell me and ask me to save the video file into inspo/ myself instead. - Node.js + Remotion: builds the graphics, captions and animations in code and renders the final video. Set up one Remotion project in this folder with a 1080x1920, 30fps composition. - Python 3 with Pillow: for contact sheets and quick checks. Run a test of each tool and show me a checklist: tool, installed yes/no, version. STEP 3: SOUND LIBRARY Fill library/sounds/ with REAL recorded sounds only, all CC0 (free to use, no credit needed), from Freesound.org or Openverse.org: - a real keyboard typing loop (2-3 options) - real mouse clicks (3-4) - soft swooshes for things sliding in (3-4) - one subtle pop for a list item appearing No cartoon dings, no synth beeps, no chimes. Write library/sounds/SOURCES.md listing each file, where it came from and its licence. If you can't download something, tell me which ones and I'll grab them. STEP 4: MOTION LIBRARY Prompt Motion (https://prompt-motion.com) is a free catalog of 200+ motion graphics made with Claude, each with the full prompt behind it. Build a small script, tools/motion.py, that: - pulls the catalog (titles, tags, preview links, prompts) into library/motion/catalog.json and refreshes it every few days - `find <tag>` shows me the best matches for a beat (tags like product-ui, diagrams, charts, phone, code, kinetic-type) - `prompt <name>` prints the full prompt so you can borrow its timing and structure Use this library for ideas on every video so the graphics don't all look the same. Borrow the idea, never copy a graphic 1:1. STEP 5: LEARN MY STYLE When the setup is done, ask me to drop 5-10 of MY best videos into my-videos/ and, only if I want a new look, 2-3 videos from creators I like into inspo/. If I add more than 10 of mine or more than 3 inspo videos, push back and help me pick the best ones. Too many references muddy the style. Then study every video: transcribe it, pull 1 frame per second into a contact sheet, and write style/style.md covering: - pacing (how often the picture changes, how long a cut holds) - captions (font, size, colour, position, how many words at a time) - layout (where my face sits, where graphics go, what stays clear for the app's buttons) - graphics (what kind: screenshots, diagrams, big numbers, app UI, logos) - colour + fonts - sound + music (when sounds hit, how loud the music sits) - hooks (what happens on screen in the first 2 seconds) Mark which choices come from my videos and which come from inspo. The goal is INSPIRED BY the inspo, not a copy. My face, my voice, my look. Show me the style guide and let me change anything before you save it. STEP 6: THE EDITING WORKFLOW (save this as a reusable skill called /edit) When I say "/edit" or "edit my video": 1. Take the newest file in raw-footage/ and ask me for the script if I haven't pasted it. 2. Transcribe it with Whisper. Find every take of every line, keep the LAST good take, cut stutters, restarts and dead air, and leave no line said twice. Show me the cut list before cutting. 3. Write a beat sheet: every sentence gets its own visual that shows exactly what I'm saying at that moment (a real screenshot, a real logo, a diagram, a number, the real app UI). Pictures over text. Check library/motion/ for ideas. Follow style/style.md. 4. Build it in Remotion: captions, graphics, sound effects on the actions (typing under typing, a click on every click, a swoosh when something slides in), and music from library/music/ under my voice. 5. Open a live preview for me BEFORE rendering anything. I'll give notes in plain words like "make the hook bigger" or "this part is boring". Fix them in the preview. 6. Only when I approve: render, master the audio to -14 LUFS so it's loud enough for Instagram, and save to exports/. 7. Before handing it over, check it yourself: watch a frame every half second and make sure nothing covers my face, nothing sits under the app's buttons, every caption matches what I say, and no part of the video sits still for more than 2 seconds. 8. Every note I give you, add it to style/feedback.md. When a note comes up twice, move it into style/style.md so you never make that mistake again. STEP 7: HOUSE RULES (put these in CLAUDE.md) - Never render before I approve the preview. - Never delete my raw footage. - Real assets beat AI-generated ones. Use real logos, real screenshots, real app UI. - Real sounds only. Nothing cartoony. - Keep all text and graphics inside the safe area (about 130px in from each side, clear of the bottom 15% where the caption and buttons sit). - Ask before using any paid tool or API. - Keep the folder lean: delete temporary renders after a video is exported. When everything is set up, print READY, then show me: the folder tree, the tool checklist, what's in the sound and motion libraries, and exactly what to do next (drop my videos into my-videos/).
2. The 4 tools it installs
You never touch these directly. But it helps to know what each one does, so you know what to ask for.
Remotion: Builds the video
Remotion makes videos out of code. That's the trick: Claude can't drag clips around a timeline, but it can write code all day. So every caption, graphic, zoom and animation gets built as code, and you watch it in a live preview before anything renders. Free for individuals and small teams.
FFmpeg: Cuts and exports
The tool almost every video app uses under the hood. Claude uses it to cut your dead air, join clips, fix the sound level and export the final file. On a Mac the prompt installs the full version, because the normal one can't read iPhone HDR footage.
Whisper: Writes your captions
OpenAI's free speech-to-text, running on your own computer. No API key, no cost per video. It gives Claude every word you said with the exact time you said it. That's how it writes captions, finds your stutters and restarts, and puts each graphic on the right word.
yt-dlp: Grabs videos to study
Downloads a video from a link so Claude can study it. You use it for the creators in your inspo folder. Instagram blocks it a lot, so if it fails, just save the video yourself and drop it in the folder.
3. The libraries most people skip
The tools let Claude edit. The libraries are what make the edit look and sound like a person made it.
- A real sound library This is why most AI edits sound cheap: the sound effects are made up beeps and dings. The prompt fills library/sounds/ with real recorded typing, real mouse clicks and soft swooshes, all CC0 (free to use anywhere, no credit needed) from Freesound and Openverse. Then the edit puts typing under typing, a click on every click and a swoosh when something slides in.
- A motion graphics library (Prompt Motion) Prompt Motion is a free catalog of 200+ motion graphics made with Claude, and every one comes with the full prompt behind it. The prompt gives Claude a little search tool for it, so before it builds a graphic it can look at what good ones do and borrow the timing and the idea. That's the difference between your graphics and the same three AI templates everyone else has.
- Your reference library my-videos/ holds 5 to 10 of your best videos. inspo/ holds 2 or 3 videos from creators whose style you want to borrow from. Claude studies both and writes your style guide. Every edit after that follows it.
4. Teach it your style
Drop 5 to 10 of your best videos into my-videos/. Pick the ones that did the best, not the ones you like the most.
Want a new look? Add 2 or 3 videos from creators you love into inspo/. Not 20. The more you give it, the blurrier your style gets.
I just added my videos. Study everything in my-videos/ and inspo/ and write my style guide in style/style.md, like the setup said. Before you write anything: - If there are more than 10 videos in my-videos/ or more than 3 in inspo/, stop and help me pick the best ones first. - Transcribe each one and pull 1 frame per second into a contact sheet so you actually see the edit, not just the words. In the guide, mark which choices come from MY videos and which come from inspo. Take ideas from the inspo, don't copy it. My face, my voice, my look. Then show me the guide as a short list (pacing, captions, layout, graphics, colour + fonts, sound, hooks) and ask me what to change before you save it.
Read the style guide it shows you and change anything that's off. This file decides how every video looks from now on, so it's worth five minutes.
5. Edit your first video
Film like you normally do. Drop the raw take into raw-footage/, then paste this with your script.
/edit Here's the script for this one: [paste your script here] Use the newest file in raw-footage/. Show me the cut list first, then the beat sheet, then open the live preview. Don't render until I say it's approved.
It cuts the dead air and the restarts, adds captions, graphics, sound and music, and opens a preview. Then you just talk to it in plain words:
- "Make the hook bigger."
- "This part is boring, show the app instead of my face."
- "The music is too loud under the second half."
When it looks right, say "approved." It renders, sets the volume for Instagram and drops the finished video in exports/. Every note you gave gets saved, so the next video starts closer to done.
The 6 things that quietly break this
None of these throw a clear error. The build prompt handles most of them for you.
- Too many reference videos This is the big one. Give it 30 videos and it averages them all into nothing. 5 to 10 of yours, 2 or 3 from other creators, max. The prompt pushes back if you go over.
- Rendering before you see it A render takes minutes. A preview is instant. If Claude renders every time you give a note, you'll wait all day. The prompt makes the live preview step mandatory and only renders once you approve.
- Your folder lives in iCloud iCloud quietly offloads files to save space. Then a render fails halfway for no clear reason. The prompt checks where your folder is and tells you to move it to your home folder if it's in iCloud.
- The normal FFmpeg on a Mac If you film on an iPhone, your footage is HDR. Plain brew install ffmpeg can't convert it, so the colours come out washed out or the export breaks. The prompt installs ffmpeg-full and checks it works.
- Instagram blocks the downloader yt-dlp often gets stopped by Instagram's login wall. Don't fight it. Save the reel to your computer and drop the file into inspo/ yourself.
- Making the same note twice Every note you give goes into style/feedback.md. When the same note comes up twice, it moves into your style guide for good. If you skip this, you'll be saying "make the captions bigger" forever.
Credit
The prompt and the workflow are mine. The tools are other people's work, given away free: Remotion, FFmpeg, Whisper by OpenAI, yt-dlp, Prompt Motion for the motion graphics catalog, and Freesound and Openverse for the sounds, which belong to the people who recorded them.