Here is the setup I use to cut raw footage into finished video, why it needs a second AI to check the work, and the two prompts that do it.
Editing needs four different abilities, and no single tool has all four. This is the split that works.
Reads the transcript and works out which take was the good one, which sentence was a false start, and where you talked to the camera instead of the audience.
Your existing subscriptionGives you the timing of every individual word, not just each sentence. Cut a fraction late and you clip the front of the next word.
Free · runs on your machineDoes the real work. Trimming, joining, audio levelling, rendering. It has run quietly behind most video software for twenty years.
Free · runs on your machineActually watches and listens to the finished cut, then reports what is wrong with it. This is the part almost nobody does.
About 10c per 13 minutesClaude reads text. It cannot watch video. So when it finishes an edit, it has no idea whether the edit actually worked.
That is worse than it sounds. A transcript will not tell you a word got clipped in half, or that the same sentence is in there twice, or that there is a two second hole in the middle. Transcription engines actively tidy those away. They are built to hand you clean readable text, so they quietly drop the stutters and the repeats.
Which means the exact faults a viewer notices most are invisible to the thing checking the work.
So you give it a second AI whose only job is to watch.
Without that loop you get an edit that reads perfectly in a transcript and sounds broken the moment you play it.
This is the part people miss. The same ability that lets it check its own work lets it study anyone else's.
Give it a video whose style you want. It watches the whole thing and comes back with the actual recipe: how long each shot is held, where the captions sit and how they animate, when the B-roll cuts in and how it moves, where the sound effects land, how the music sits under the voice.
That breakdown becomes a spec. Then it edits your footage to that spec.
So instead of describing the look you want and hoping, you hand over an example and get the same treatment applied to your own material. Reverse engineering a style used to take an editor a day of frame-by-frame work. It takes about a minute.
It is describing what it observes, not copying the footage. You get the editing grammar: pacing, caption behaviour, cut rhythm. The content stays yours.
You do not have to paste anything below. Claude Code can read this page itself. Open it, paste the line underneath, and it will do the whole setup for you.
Read https://ai-video-editor.asmaldigital.com and set this up on my machine. Follow it exactly. Check what I already have first, ask me before installing anything, and prove each piece actually works before moving on. Then tell me in plain English what each part does. When the setup is done, stop and wait. Do not start editing a video until I give you the file.
Some setups cannot fetch a web page. If Claude tells you it cannot read the URL, use the two paste blocks further down instead. They contain exactly the same instructions, just handed over directly.
You do not need to install anything by hand. This makes Claude work out what you already have and set up what you don't.
The prompt is written to ask you before it installs anything. Never let any AI install software on your machine without telling you first. Including this one. If it starts installing without asking, stop it.
I want to set up an AI video editing workflow on this machine. Walk me through it one step at a time, in plain English. I am not a developer. STEP 1 - Check what I already have. Tell me which of these are present and which are missing: - ffmpeg and ffprobe - a local speech-to-text tool that gives word-level timestamps - whether GEMINI_API_KEY exists in my environment STEP 2 - For anything missing, DO NOT install it yet. For each one tell me: - what it is, in one sentence - what job it does in video editing - roughly how big the download is - the exact command you would run Then stop and wait for me to say yes. Ask about one thing at a time. STEP 3 - Once each piece is in, prove it works by running it on a tiny test file you create yourself. Do not tell me something works until you have actually run it. STEP 4 - Tell me in plain English what each piece does. RULES: Ask before installing anything. Do not change system settings. Do not use sudo without asking. If something fails, stop and tell me what failed and why. Do not silently try another approach.
Fill in your file paths at the top. Everything from step 9 down is the part that makes the result publishable, so do not cut it out.
Edit this recording for me.
FILES
camera: <path to your camera file>
screen recording (leave out if none): <path>
PART 1 - WORK OUT THE CUTS
1. Pull the audio out and transcribe it with word-level timestamps.
2. Find every silence longer than about 0.8 seconds.
3. Find every place I said the same line more than once trying to get
it right.
4. For repeated lines, keep the LAST complete take and drop the earlier
attempts. The last one is almost always the good one.
5. Also cut anything where I talk to the camera as an editor, like
"leave that bit out" or "let me start again". Cut the instruction as
well as the part it refers to.
6. Never place a cut inside a word. Use the word-level timings so every
cut lands in the silence between words.
7. Where the gap between two kept sections is under about 0.8 seconds,
leave it. That is me breathing, not dead air.
SHOW ME THE CUT LIST BEFORE YOU RENDER. I want to approve it.
PART 2 - BUILD IT
8. Render the edit. Keep the audio at a consistent level throughout.
PART 3 - CHECK YOUR OWN WORK (do not skip this)
9. Split the finished video into overlapping one minute chunks.
10. Send each chunk to Gemini and ask it to LISTEN and report:
- any word cut off part way through
- any sentence that appears twice
- any silence longer than a second
- anything that sounds abrupt or wrong
11. Fix what it reports, at those exact timestamps, and re-render.
12. Run the check again on the parts you changed.
13. Repeat until the check comes back with nothing.
Do not tell me the video is done based on the transcript. The transcript
will look fine even when the audio is broken. Only the listening check
counts.
Add this to the end of Part 1: "Here is a video whose style I want. Watch it and give me the shot lengths, caption style and cut rhythm, then apply that to my footage."
Not the cutting. Brief someone, wait, review, send notes, wait again. This collapses that to one person and one instruction.
See what else I build