Automated Video Editing with Claude Code: My Real Stack

Answer in 60 seconds. Claude Code can drive FFmpeg, Remotion, OBS (the free screen-recording app), and the big editors (Premiere, DaVinci Resolve, Final Cut Pro) through MCP, the protocol that lets Claude talk to other apps. You write a sentence and Claude writes the actual command. Your machine spits out the file. The parts I run weekly: FFmpeg for resize, crop, compress, and convert. Remotion for Shorts that build themselves from a small data file. Planning Mode plus a CLAUDE.md file for templates I reuse. The pieces I have not yet pushed through a paid edit: Jumper-style links into Adobe and Resolve, Hyperframes for motion graphics, video-use for clips that clean themselves up, and OBS scripts for live captures. The pieces I run cost about $0.04 to $0.20 per finished clip in running costs. A freelancer charges $80 to $300 for the same cut.

Resize a clip for Instagram square, again for TikTok. Then again for X. Three exports out of CapCut, the wrong one always to the wrong folder of course…and the whole loop takes an hour at minimum, every time you publish.

That 17-second motion reel above? The source is a NotebookLM explainer I made about agentic video editing. The shorter, punchier version with the pan-and-zoom and cross-dissolves was built by Claude Code using FFmpeg, picking the strongest 6 moments from the original 2:17 and stitching them with cross-fades. I dropped the source in this project folder and wanted an Instagram square and a Shorts vertical of that reel. I typed two sentences in Claude Code:

Resize this clip to 1080x1080 for Instagram, center-crop so nothing important gets squished, keep the audio.

Then:

Now do a 9:16 version for Shorts and TikTok.

Claude wrote the FFmpeg commands, ran them on my machine, dropped both files in static/videos/agentic-video-editing/. Under a minute. The two variants you’ll see further down are the real outputs from those two prompts on the exact video above. The hour you used to lose? Forget about it!

Quick truth before I sell you anything: I am not a super-skilled video editor. I am quite capable but I do not cut documentaries for a living. What I do is publish some videos on the side of writing, and the resize-and-export part of that loop was eating my time. The stack below is built around that one shape of work. If you cut for clients all day, the parts you need are different, and Section IV is where you should start, not Section II.

This piece is the working stack behind that one minute, plus the parts I have read about but have not run myself. Honest sizing on both columns. If a section says “I haven’t tried this yet”, that is on purpose. The honest instead b.s. rules in this house do not let me pretend.

Key Takeaway: Agentic video editing means Claude Code runs a bunch of small tools (FFmpeg, Remotion, MCP servers) one after another. You write a sentence. Claude does the clicking.

What “agentic editing” actually means here

Agentic is one of those words AI folks throw around to sound smart. Most of the time it means you want to be fancy and vague. Here it means one specific thing: Claude Code reads your sentence, decides which commands to run in what order, runs them, looks at the output, and fixes it if something went wrong. You did not click anything. You wrote one sentence.

Think of it like hiring a junior video editor who follows orders, never gets bored, never asks for overtime, and only knows the tools you point at. You say “make me a 1080x1080 version with the audio kept.” The junior figures out which command does that, runs it, and hands you the file. He has no taste. You still bring that.

Old way: you drag clips around for hours. New way: you type a sentence and Claude does the dragging. The swap works for the 80% of jobs that are the same boring cuts every time. The 20% where taste matters still wants a real editor. So you mix the two: AI for the boring stuff (it’s meant for this, you know), you for the calls that matter more.

FFmpeg without the flag-stuffed mess (the part I run weekly)

FFmpeg is a free video tool that runs behind most of the video on the internet. It is also famous for commands packed with weird flags nobody remembers. YouTube tutorials about FFmpeg syntax have the most “rewatching this for the 50th time” comments of any topic in tech.

Claude Code hides the messy flags for you. You describe the result in English, Claude writes the command, you read the command before running it, your machine spits out the file.

Why this works. Claude has read more FFmpeg docs and Stack Overflow threads than you ever will. What takes you 30 minutes of Googling takes Claude 2 seconds. Your job is to read what Claude wrote before you hit enter.

Here’s the actual prompt I sent:

Resize this clip to 1080x1080 for Instagram, center-crop so nothing important gets squished, keep the audio.

Claude wrote this command:

ffmpeg -i Agentic_Video_Editing.mp4 \
  -vf "scale=1080:1080:force_original_aspect_ratio=increase,crop=1080:1080" \
  -c:a copy instagram-square.mp4

It ran in a few seconds. The 1080x1080 square version saved next to the source:

Same source clip, second prompt:

Now do a 9:16 version for Shorts and TikTok.

Claude wrote a slightly different crop (1080x1920 instead of 1080x1080), ran it, and the vertical version saved alongside:

Two prompts, two real outputs from the same source. No timeline, no scrubbing.

The boring tasks Claude is good at:

• Resize and crop for Instagram square, Reels, Shorts, TikTok
• Compress for under-25MB email attachment limits
• Convert MOV to MP4 because your blog or CMS (the back-end of your website, like WordPress) won’t accept MOV
• Strip audio out for transcription
• Speed up b-roll 2x to fit a 30-second slot
• Add a padded white border around a clip with a QR code in the corner

If you have ever lost two hours to any of those tasks, this is where you start. Open a terminal (the black-screen window where you type commands instead of clicking), start Claude Code, point it at the folder with your source files, and send the prompts one at a time.

Remotion and React (the part where Claude writes code so you don’t have to)

Remotion is a free tool that lets you build videos with code. You write a small bit of code in React and TypeScript (the same languages most modern websites are built in), Remotion turns the code into an MP4. The trick: Claude can write that code for you while you describe the video.

I set up a Remotion project in this project folder last week with one command, npx create-video@latest. (npx is a small helper that ships with Node.js and runs the setup for you, no real install.) Claude wrote a Composition (Remotion’s word for a video template) that builds a top-five listicle from a small data file. That data file is just plain text holding a list of items with their names, durations, and colors. Title cards and animated bars run over a brand-dark background, with music ducking under the voice-over. The data file holds the five items, their durations, the brand colors, and the music timing. Change the data file, hit re-render, you get a different video. The Composition stays the same.

Why this matters for non-developers. You do not learn Remotion, instead you learn the prompts that make Claude build Remotion components for you. Once one Composition exists in the project, every future video that fits that template is a small change to the data file.

Why it pays off. Videos built by code are easy to repeat. The same template runs forty videos this year and forty next year for the cost of editing one data file. If you publish the same kind of explainer every week (same intro, same lower-third (the graphic at the bottom of the screen with someone’s name), same caption font, same outro), code is the cheaper long game. If every video looks different, the timeline still wins.

Where I would not use Remotion: anything where the camera moves for an effect, anything cut to a feeling, anything that needs to look like a movie. That is human-editor work. Remotion is for the explainer template, the data-viz reel, the Shorts version of your blog post, and any other format you publish on a schedule.

Key Takeaway: Remotion turns a video format you publish on repeat into a small data file. You stop cutting the same intro forty times a year.

Jumper, MCP, and the big editors (I haven’t tried this myself)

This is the part the AI video fanatics have been talking lately. The pitch: Claude Code connects to Premiere Pro, DaVinci Resolve, or Final Cut Pro through a Model Context Protocol connector (the most cited one is called Jumper). You ask Claude to find every clip where the guest smiles, and Claude searches the transcripts and visual data and pulls out those clips. Same idea for “every B-roll shot of the kitchen” (B-roll is the supporting footage you cut to between your main shots) or “all the moments the speaker said bullshit.”

Jumper has not seen any of my paid edits yet. I have read the manuals and the case studies, mostly from documentary editors who say it cut their assembly time (the rough first arrangement of clips) in half on long jobs.

What I would check before trusting it on a paid edit:

1. Where the media goes. Jumper’s documentation says media stays on your machine and only the descriptions move through Claude. Read the data flow for your specific tool before you point it at footage you signed a confidentiality agreement to protect.
2. The visual search quality. “All moments the guest smiles” sounds great until the model misses small smiles and over-flags the host. I would test it on 30 minutes of footage I already know before I told a client this would save them time.
3. The undo path. Claude Code can do real damage to a Premiere project file. The MCP connector should never touch the original timeline. Work in a duplicate.

Treat this as a research note. Once I have run my first cut through one of these connectors, that piece earns its own article with screenshots and timing.

Hyperframes and motion graphics (also untried by me)

Hyperframes is an online tool that builds motion graphics from a sentence. You type “slide a lower-third in from the left with the speaker’s name and title, hold for 5 seconds, slide out,” and the tool produces the clip on its servers. Same idea for logo reveals, end cards, and Shorts where text flies around the screen.

The promise is that you skip the timeline entirely. No fiddling with animation, no After Effects (the Adobe app most motion-graphics work happens in). The trade-off is that the rendering happens on their servers. You need an account, you wait in a queue, and you pay per render.

The pieces of the pitch I am skeptical about: does the animation feel cheap or look like real taste, does the same prompt give you the same look twice, and what does it actually cost when you use it a lot.

If I tested it, my first run would be the same lower-third I currently do in CapCut, side by side, same fonts, same colors, same duration. Compare the two. If Hyperframes wins on time and the look does not feel fake, it earns a slot. If it loses, you stay on the timeline you know.

video-use and the self-cleaning clip (research, not yet receipts)

video-use is a framework that “watches” a clip by reading the transcript and a few summary images, instead of looking at every frame. The headline feature is the self-running polish loop: Claude scans the clip, removes filler words (“um,” “uh,” “you know”), adds a color grade (the colour-and-contrast pass that makes everything match), drops 30ms audio fades at every cut to stop the clicking sound, and checks the output at every cut for visual jumps.

Nothing real has gone through video-use on my machine yet. The idea makes sense on paper. The risk is the same as any “fully self-running” pitch in 2026: you trust it once, it eats a key clip, and you lose a full day fixing it.

What I would test on the first real run:

• The filler-word remover on a podcast clip where I already know the count of “ums.” Does the cleaned version match what I would cut.
• The color grade on a one-camera interview shot in mixed lighting. Does it keep skin tones natural.
• The cut check on a fast-cut sequence. Does it flag a real visual jump or invent one.

That comparison is its own piece, when the receipts exist.

Planning Mode and CLAUDE.md (the part most readers skip)

This is the section most people scroll past. It is also the one that prevents the disaster.

Planning Mode is a Claude Code mode you reach by pressing Shift+Tab twice inside the Claude Code terminal. The first press puts you in auto-accept-edits mode, the second press puts you in Plan Mode, and the status bar at the bottom shows which one is active. In Plan Mode, Claude writes you the plan first instead of running commands. You read it, fix anything wrong, then approve. For video work this is the difference between “Claude rewrote my source file because of a bad path” and “Claude described what it was about to do and I caught the bad path.”

I use Planning Mode every time the operation touches the original media file. Re-encoding, batch processing, anything that overwrites. Two minutes of reading the plan saves a half-day of recovery work.

CLAUDE.md is a plain text file in the top of your project folder that Claude Code reads at the start of every session. The .md is for Markdown, which is text with simple formatting like bold and headings. Anyone can open it in any text editor. Think of it as the project’s standing instructions. For video work it stores the brand colors and fonts for lower-thirds, the default output sizes, the FFmpeg flags I always want, where finished files go, and the names of the templates I reuse. Claude reads CLAUDE.md and behaves the same way every time you open the project.

Side benefit: the file is portable. Drop the same CLAUDE.md into a new project and Claude gets the same setup.

For OBS, the same idea works. You write a small script Claude can call (switch a scene, mute the mic, apply a rounded-corner filter, fade in a music bed), and you put the script’s name in CLAUDE.md so Claude knows it exists. I have done this for screen-recording sessions. The OBS script runs once per session and saves you all the clicking.

Key Takeaway: Planning Mode protects you from one bad command. CLAUDE.md gives you the same setup every time. The two together turn Claude Code from a clever toy into a working tool.

Cost versus performance: the math

Cost numbers I pulled from the posted AI prices (each company charges per million words of work, called tokens), on a 1-minute 1080p clip:

Numbers are rough. Your usage will swing depending on how much planning Claude does and how many retries it takes. Plan on $0.04 to $0.20 per finished clip as the real range for the FFmpeg-and-Remotion stack.

The freelancer comparison: a basic resize-and-subtitle pass on Fiverr or Upwork runs $30 to $80 per video. A real human editor cutting a one-minute clip runs $80 to $300 depending on the city. The math only matters if you publish more than one of these a week. At one-off rates the human still wins on taste.

Mix the two like this: AI for the cuts you do over and over (resize work, subtitle burns, format swaps, that 30-second X pull). You for the cuts a client will feel (which take of a line lands, which beat the music drops on, which version looks like you and not a template). Same idea as my 3-tool AI stack: pick the right tool for each part of the job. Same play in the NotebookLM SEO workflow: AI does the boring 80%, you handle the 20% that decides if it lands.

Recordable steps for the demo session

This part is for me, for next session. The article above will pair with a real screen-recording demo, recorded against the actual workflow, not fake mockups. Sections I can demo on my machine right now:

• Section II (FFmpeg). Open a terminal in this project. Show the source Recording - April 30, 2026.mp4. Type the resize prompt. Show Claude writing the FFmpeg command. Run it. Open the resulting instagram-square.mp4 in the file explorer. Repeat for the 1080x1920 Shorts version.
• Section III (Remotion). Open tools/screen-demo/. Run npm run preview (Remotion opens a preview at a local web address only your computer can see). Show the data file driving the listicle. Edit the data file. Re-render in the preview. Open the finished MP4.
• Section VII (Planning Mode + CLAUDE.md). Open Claude Code in this project. Hit Shift+Tab to enter Planning Mode. Ask Claude to do a multi-step FFmpeg job (trim, resize, burn subtitle, export). Show the written plan. Approve. Show the resulting clip.

Sections IV, V, VI stay text-only in the article until I have tried them on real footage. No fake demos.

Recording target: 60 to 90 seconds per section, captured with the local recorder at tools/recorder/, branded with the AI Stash overlay. Outputs per run: 1920x1080 hero, 1080x1920 Shorts, 960px GIF for inline blog use, plus a script.txt for the LinkedIn caption.

FAQ

Does Claude Code require coding knowledge for video editing?

No. You tell Claude Code what you want and it writes the FFmpeg command, the Remotion code, or the MCP call for you. What you do need is a terminal, Claude Code installed, and the patience to read what it wrote before you run it. Reading the output is the only place a non-coder still has to pay attention.

How long does Claude Code take to process a video?

For everyday jobs like resizing, compressing, or trimming a 1080p clip up to 2 minutes long, the FFmpeg command itself runs in 5 to 30 seconds. The slow part is Claude reading the request and writing the command, which adds another 10 to 20 seconds. Remotion takes longer because it builds the video frame by frame, usually 1 to 4 minutes for a 60-second 1080p clip.

Can Claude Code work with CapCut or iMovie?

Not directly. CapCut and iMovie do not give Claude a way to drive them from outside. Claude works as a standalone editor calling FFmpeg or Remotion, or it links into the big editors (Premiere Pro, DaVinci Resolve, Final Cut Pro) through Model Context Protocol connectors. If you live inside CapCut, Claude is not the right tool for that work.

Is my footage uploaded to the cloud when Claude Code edits it?

It depends on the tool. FFmpeg runs entirely on your machine. Tools like Jumper that link MCP into Premiere or Resolve are built to keep media on your machine and only send the descriptions and commands to Claude. Remotion can render on your machine or on a cloud farm, your choice. Hyperframes and other online tools do need cloud rendering. Always read the data flow for each tool before pointing it at client footage.

Can Claude Code generate vertical 9:16 videos for TikTok and Shorts?

Yes. You ask for a 1080x1920 output, Claude writes either an FFmpeg crop or a Remotion file with the right size, and you get a Shorts-ready clip. Subtitle styling per platform (different fonts and positions for TikTok versus YouTube Shorts) can be set in the same prompt. I do this every week for the same source clip and it costs cents.

What I would do this week if I were you

If you publish video weekly and you currently spend more than two hours per week on resize-and-export work, install Claude Code, point it at a folder with three of last week’s source files, and ask it to produce all three platform versions. The prompts that work for video are the same shape as the ones I use for everything else, see how to use Claude Opus 4.7 for the basics of how I prompt. That is your first real receipt. If the receipt lands, the FFmpeg layer of this stack pays for itself in the first month.

If you publish less than weekly, the Remotion layer is where I would start. Build one Composition for the format you repeat the most. After that, every clip after the first comes from a quick data-file edit. The savings only stack up if the template exists.

Three pieces stay parked until they earn their receipts on a paid cut: MCP connectors into the big editors, Hyperframes, video-use. Each gets its own write-up the day I have screenshots. Subscribe to the Stash if you want them when they land, and the working examples when they break.