Skip to content
TMTiger Mendez

GuidesOctober 27, 2026

Step-by-step guide · Claude

Your video editor is Claude.

Works on Mac and Windows. We show you the steps for your computer.

Record in one take, drop the video in a folder and Claude does the cuts, the captions, the zooms and the music in your style. No paying editors, no learning CapCut. This is the system I use for my reels.

Raw
Edited
Subs
Ididn'teditthisreelmyself
Audio
silenceretakeumm
Claude listens to the video word by word, removes what's extra and puts it back together. In this example, 12 of 55 seconds were silences, filler words and retakes.

00:00Before you start

What you need

  • A computer

    A Mac on macOS 13 or newer. 16 GB of RAM is better if you'll make animated graphics.A PC on Windows 10 or 11. 16 GB of RAM is better; with an NVIDIA card, transcription and cutouts run faster.

  • Claude in the desktop app

    With Claude Code or Cowork, which work with the files on your computer. You need a paid Claude plan.

  • Your video, vertical

    Shot on your phone, at 1080×1920 or in 4K.

  • 3 to 5 videos you like

    Yours or from creators you admire. That's where your style comes from.

You don't need to know how to code. Claude writes and runs everything; you approve and give feedback, like you would with a real editor.

00:12Step 1

Set up the folder

Claude works inside a folder. If everything has its place, it never touches what it shouldn't and always knows where to look.

ai-editor/

raw/# your videos straight from the phone. Claude never changes them

references/# the videos you like, to pull out your style

output/# the plans and finished videos: reel_v1.mp4, reel_v2.mp4…

style.md# your style guide (Claude writes it in step 3)

CLAUDE.md# the house rules. Claude Code reads it at the start of every chat

  1. 1Create the ai-editor folder with its three subfolders. In Finder, inside Documents (Cmd + Shift + N makes a new folder).Create the ai-editor folder with its three subfolders. In File Explorer, inside Documents (Ctrl + Shift + N makes a new folder).
  2. 2Open Claude in the desktop app, pick Claude Code (or Cowork) and give it access to ai-editor.
  3. 3Ask it: “Create a CLAUDE.md saying you're my video editor, that you read style.md before editing, and that you never change anything inside raw/ or references/.”

TrickThe CLAUDE.md file is the difference between repeating your instructions in every chat and Claude already knowing how you work. Anything you correct twice, add it there.

00:24Step 2

Give it the tools

Claude doesn't edit by magic: it uses the same free programs studios use. They run on your computer, so your videos aren't uploaded anywhere.

  • ffmpeg

    Required

    Cutting, joining, detecting silences, leveling the audio and exporting the final video. ffmpeg.org

  • Whisper

    Required

    Listening to your video and writing down every word with the exact second you say it. The cuts and captions come from there. github.com/ggml-org/whisper.cpp

  • Remotion

    Required

    Animated captions, cards, counters, zooms and graphics made with code. Free for individuals and small teams. remotion.dev

  • HyperFrames

    Required

    Graphics built like web pages, sound effects and cutting you out of the background to put text behind your head. github.com/heygen-com/hyperframes

  • yt-dlp

    Required

    Downloading your reference videos from an Instagram, TikTok or YouTube link. github.com/yt-dlp/yt-dlp

  • DeepFilterNet

    Optional

    Removing background noise and echo from your voice when you record at home. github.com/Rikorose/DeepFilterNet

Prompt · install · Mac
I want you to be my video editor. You work in this folder (ai-editor). My computer is a Mac.

Install and test these free tools. If one is already there, just check it:
1. Homebrew, if I don't have it (brew.sh).
2. With Homebrew: ffmpeg, yt-dlp, whisper-cpp, node and python@3.11.
3. The Whisper large-v3-turbo model (or small if my Mac is slow) from huggingface.co/ggerganov/whisper.cpp, in a models folder.
4. Inside ai-editor: HyperFrames and Remotion with npm, plus their Claude skills (npx skills add heygen-com/hyperframes and npx skills add remotion-dev/skills).
5. (Optional) DeepFilterNet with Python 3.11, to clean up noise in my voice.

Don't use anything paid or anything that asks for an account or an API key.
Tell me in one line what you're going to install before you do it.

When you're done, run a test: make a 3-second clip with some text and a tone, transcribe a short audio, cut a person out of a photo with HyperFrames, and tell me what worked and what didn't.
Prompt · install · Windows
I want you to be my video editor. You work in this folder (ai-editor). My computer runs Windows.

Install and test these free tools. If one is already there, just check it:
1. With winget, from PowerShell: Gyan.FFmpeg, yt-dlp.yt-dlp, OpenJS.NodeJS.LTS and Python.Python.3.11.
2. whisper.cpp: the Windows zip (whisper-bin-x64.zip) from github.com/ggml-org/whisper.cpp/releases. If my PC has an NVIDIA card, the CUDA build. Add it to the PATH.
3. The Whisper large-v3-turbo model (or small if my PC is slow) from huggingface.co/ggerganov/whisper.cpp, in a models folder.
4. Inside ai-editor: HyperFrames and Remotion with npm, plus their Claude skills (npx skills add heygen-com/hyperframes and npx skills add remotion-dev/skills).
5. (Optional) DeepFilterNet with Python 3.11, to clean up noise in my voice.

Use Windows paths and PowerShell. If a command needs JSON, save it to a file instead of typing it on the command line.
Don't use anything paid or anything that asks for an account or an API key.
Tell me in one line what you're going to install before you do it.

When you're done, run a test: make a 3-second clip with some text and a tone, transcribe a short audio, cut a person out of a photo with HyperFrames, and tell me what worked and what didn't.

NormalClaude will ask permission to install things and run commands. Read the line it shows you and accept. Your Mac may ask for your password to install Homebrew: type it yourself in the system window. If something fails, paste it the error and let it fix it.Claude will ask permission to install things and run commands. Read the line it shows you and accept. Windows may ask whether to allow changes to your device: say yes to winget and Node. If something fails, paste it the error and let it fix it.

00:36Step 3

Teach it your style

Here's the part almost nobody gets right. Claude doesn't “watch” a video like you do: it turns it into frames, text and numbers. It makes contact sheets, counts the cuts, measures how long each shot lasts and transcribes what's said. With that it writes your style guide.

Prompt · get my style
[WHAT I WANT]
Study the videos in references/ and write my style guide in style.md.
If I give you links, download them first with yt-dlp into references/. They're only for studying the style: never to repost them.

[HOW TO ANALYZE THEM]
For each video:
1. Make a contact sheet (1 frame every 2 seconds) and another of the first 3 seconds (2 per second). Look at them.
2. Count the cuts with ffmpeg's scene detection and work out the average shot length.
3. Transcribe it with Whisper with word-level timestamps.
4. Note down: the hook (what you see and hear at second 0), the captions (size, position, color, highlighted words, animation), the palette with hex codes, the graphics and when they appear, the music and sound effects, the framing and the call to action.

[WHAT TO DELIVER]
1. A short table: one video per row with length, cuts, hook and its strongest device.
2. What repeats in at least 2 videos. That's style; what shows up in only one is chance.
3. style.md with: the essence in one sentence, hook, pacing, captions, color, fonts (with free Google Fonts alternatives), graphics, sound, framing and what I DON'T do.

[RULES]
- Copy the visual language, never other creators' lines, brands or content.
- If something can't be known for sure (like the exact font), say so and suggest the closest one.
- Colors and timings measured, not made up.
- Show me the guide and ask what to adjust before saving it.

What a style.md looks like (an example, based on mine)

Template · style.md
# My style

Essence: an engineer who explains it to you like a friend. Clear, direct, no fluff.

Hook: the whole idea in frame 0, in big text at the top. 6 words max.
Pacing: no silences longer than 0.3 s. A gentle zoom every 4–6 s if the shot doesn't change.
Captions: word by word, centered at mid height. Never in the bottom 20 % (Instagram's buttons go there). The key word in the accent color.
Color: dark background #0d0f13, text #ffffff, accent #ff6a33.
Fonts: titles in Bricolage Grotesque 800, captions in Instrument Sans 600.
Graphics: real screenshots, never mockups. They come in with a short slide and a soft sound.
Sound: voice at −14 LUFS. Instrumental music at 10–15 %. A "whoosh" only when a graphic comes in.
Framing: face in the top third. Nothing covers it.
Ending: one word to comment, big and in quotes.

## What I DON'T do
- Emojis on the cover or in on-screen text.
- Over-the-top transitions (spins, cubes, flashes).
- My @handle inside the video (Instagram already shows it).
- Facts or figures I didn't say.

00:42Step 4

Make it yours

Before editing, give it your colors and your font. Almost everything can be customized: the color and size of the text, how the captions look and where they go, when it zooms and how hard, the transitions, the music and the sound effects. If you say nothing, it still gets good results; if you tell it, it comes out the way you pictured it.

Your colors
Captions
Zoom
Sound
Prompt · my brand (built from your choices)
Before editing, save this in style.md as my brand and use it in all my videos:

- Colors: main text #ffffff, accent #ff6a1a (for the key word and details), graphics background #0d0f13.
- Font: Montserrat (from Google Fonts) for titles and captions.
- Captions: 1 word at a time, at mid height, never in the bottom 20 %. The key word in the accent color.
- Zoom: a gentle zoom every 4–6 seconds if the shot doesn't change.
- Sound: instrumental music at 10–15 % under the voice, and a short sound effect on every animation. The voice always on top.
- Nothing covers my face. No emojis in the big text.

If any of this can't be done exactly, tell me what the closest option is.

Tips that make the difference

Use a reference for each video

Paste the link to a reel you like and tell it what to take from it.

“I want the pacing of this reel, but with my colors.”

Call the animation by its name

Text behind me, punch-in zoom, sticker cutout, split screen, counter.

“When I say “free”, text behind me.”

Build your sound folder

Keep your effects in a sounds/ folder. There are sites with free effects, like pixabay.com/sound-effects (commercial use, no attribution). Ask Claude for a card for each one: how long it lasts and what it's for. That way it picks the right sound without you saying so.

“Create sounds/manifest.json with a description of each sound.”

Give the exact second

The more precise, the fewer rounds.

“At second 8, punch in on my face.”

Show it a screenshot

Drag an image of how you want the caption or title into the chat.

“Make the captions like this, in my color.”

Ask for options

For the hook or the cover, have it suggest 3 and you pick.

“Give me 3 hooks under 6 words.”

Give it your logo or colors as an image

Claude pulls the color codes from a photo or your logo.

“Get my palette from this image.”

Save what you like

That way your next video comes out right from the first version.

“Add this to style.md.”

00:48Step 5

Record and let it edit

Record with the edit in mind, drop the clips in raw/ and paste the master prompt. If you slip up, repeat the whole sentence and keep going: Claude keeps the best take.

Record like thisPhone on a stand, space above your head and a short pause between sentences: that way the cuts and effects fit. If you slip up, clap before repeating: Claude sees it in the audio and knows where each attempt starts.

From your phone to your computerFrom an iPhone, with AirDrop (it lands in Downloads). From Android, with a USB cable or by uploading it to Google Drive. Send it in original quality, not over WhatsApp: WhatsApp compresses it.With a USB cable (the phone shows up in File Explorer), with the Phone Link app or by uploading it to Google Drive. Send it in original quality, not over WhatsApp: WhatsApp compresses it.

  1. 1 · Claude

    Transcribes

    Every word with its timestamp.

  2. 2 · Claude

    Finds what's extra

    Silences, filler words and retakes.

  3. 3 · Claude

    Proposes a plan

    Scene by scene, with every effect.

  4. 4 · You

    You approve

    Or ask for changes before it builds.

  5. 5 · Claude

    Builds and exports

    Cuts, captions, graphics and sound.

  6. 6 · Claude

    Checks itself

    Looks at key frames and measures the audio.

Prompt · edit
[MY VIDEO]
- File: [name in raw/, or "the most recent"]
- What it's about and who it's for: [topic + audience]
- Max length: [e.g. 40 seconds]
- Ending: [e.g. "Comment EDITOR"]
- Tweaks for this video only: [e.g. "no music", "more zooms"] (optional)

[HOW TO WORK]
1. Read style.md and my brand (colors, font, captions, zoom and sound).
2. Transcribe the video with Whisper, with word-level timestamps. Detect silences with ffmpeg and mark filler words and retakes (if I repeat a sentence, keep the last one).
3. If something's missing, ask me up to 4 multiple-choice questions with your recommendation marked.
4. Save a scene-by-scene table in output/plan.md: time, what I say, what's on screen, which effect it gets. Wait for my approval.
5. Write the edit as a script (edit.py or a Remotion project) in output/, so each fix means changing a line and exporting again, not starting over.
6. Export as MP4 1080×1920 at 30 fps, with the voice at −14 LUFS and the music underneath.
7. Before delivering, grab frames at the hook, at each graphic and at the ending, and look at them. Check that nothing covers my face, that the captions are in sync with the voice and that there are no typos.
8. Save it as output/[name]_v1.mp4. Each fix is a new version (v2, v3); don't delete the old ones.

[RULES]
- Never change or delete anything in raw/ or references/.
- Don't make up facts or figures. If something's missing, ask me.
- Don't add logos, brands or material that isn't mine.
- Don't use paid services or outside AI without asking me.

[DELIVERY]
The video, the plan and a short summary: which effect goes where and what's easy to adjust.

Why a scriptIf Claude edits “by hand”, every change can break something else. If the edit is a script, telling it “cut the pause at second 12” means changing a number and exporting again. That's how I edit my reels.

01:00Step 6

Give notes like you would to an editor

The first version is almost never the final one. Tell it what to change, where and at which second. The more specific, the faster.

  • 00:00–00:02“Make the hook bigger and have it land with a hit.”
  • 00:12“Cut the pause before “and then”.”
  • 00:18–00:24“Move the captions up, they cover my hand.”
  • all“Turn the music down, it's drowning my voice.”
  • 00:27“When I say “four tools”, show all four.”
  • always“Add to style.md that I never use spin transitions.”

What nobody tells youNot everything comes out the way you want on the first try, and that's fine. Fix it in the same Claude chat or finish it in the editor you like (CapCut, Edits, Premiere). Claude saves you the heavy lifting; the final touch can be yours.

01:12Use less

Edit faster and use less

How much of your plan does it use? There's no fixed number: it depends on how many times you ask for changes. Editing video is one of the heaviest uses, because in every round Claude reviews everything again. On the Pro plan you can do several reels a week if you follow these five tips. I learned them the hard way: the reel for this guide took me 16 versions because I kept asking for things as I went.

  1. 1

    Give it everything from the start

    The script, the style, the effects and the sounds, in the first message. Everything you decide later is one more round.

  2. 2

    Only give it the good takes

    Trim them on your phone (Photos on iPhone, Gallery on Android) or tell it the seconds of each one. That way it doesn't transcribe or compare repeated attempts, and it makes fewer mistakes.

  3. 3

    Put all your changes in one message

    A list of everything you want changed, each with its second. Fewer rounds, less usage.

  4. 4

    Sonnet to edit, Opus for the plan

    Opus thinks the plan through better; Sonnet edits just as well and uses less. Switch it in the model picker.

  5. 5

    A new chat for each video

    A long chat drags everything before it along and every message costs more. Whatever it should remember, save it in style.md, not in the chat.

01:24Before you post

Check these 8 things

0 of 8 checked

01:36Bonus

5 hooks that work

I studied the “I edit with Claude” reels with the most comments. They all open with one of these five formulas. Swap in your topic and use it.

“I didn't edit this reel.”

The claim. You say the result at second 0 and people stay to see if it's true.

“I gave it my raw video. Look what it gave me back.”

The before and after. Split screen: the raw on one side, the edit on the other.

“I used to spend 5 hours editing every reel.”

The pain, with a number. It only works if the number is yours and it's true.

“Everyone's editing with AI. Here's what they don't show you.”

The trend, with something new. You ride what people are already watching and promise the missing part.

“Watch it cut this filler word… umm… this one.”

The live proof. You say the filler word on purpose and the cut removes it in front of everyone.

Now record yours

And if you want me to set up this system, or another one with AI, for your business or your team, let's talk for 30 minutes.

If you make your reel with this guide, tag me on Instagram @tigermendez_ and I'll share it.

More guides

See all guides