Every time I asked an AI to edit one of my reels, I explained the whole process again. And even with the explanation, it changed the style from one video to the next, because there was no defined workflow. So I stopped teaching the same thing every time: I turned each step of the edit into a skill.

A skill is a file of instructions you install in the AI agent. The agent reads the skill on its own when your request matches it and follows the process written there, instead of inventing a new one for every video. You teach it once.

Inside, a skill is text. The description at the top tells the agent when to use the skill, and the rest is the process, step by step:

The page of the SKILL.md file of the reel-cut skill on GitHub, with the name, the description highlighted in yellow and the beta status.
The start of the cutting skill: the name, the description that says when to use it and the status.
Important

The reel skills run in Claude Code, which also lives inside the Claude desktop app, in the Code tab. The regular chat can’t run them, because they use programs and files on your computer.

The top of the Claude sidebar, with the Code button next to the chat button and the “Code” tooltip showing.
In the Claude desktop app, the Code button sits at the top of the sidebar, next to the chat.

These are the skills, in the order they come into a reel:

Skill What it does What you approve
reel-script Writes the script from your notes, with a hook, one closing call to action and the teleprompter text timed the script, before recording
reel-cut Joins the recordings, picks the best take of each line and removes the pauses the takes table, before the cut
reel-motion Decides which parts become animation and builds the scenes timed to each word the plan and the still frames, before rendering
reel-finish Assembles the reel as a CapCut project or as a finished mp4, in the style of one of your projects the assembled reel
post-caption Writes the post caption the caption
reel-edit Runs cut, animation and finishing in order, without you calling each one each step, before the next

In the first line of a reel of mine about the electronic voting machine, which debunks that rumor, I recorded the same line four times. The cutting skill kept one:

Four frames of the same line, “As urnas eletrônicas são fraudulentas.” (“Electronic voting machines are rigged.”), recorded in three files. The first three are faded, with the notes low rehearsal voice, repeated in the same file and followed by “De novo… corta aí” (“Again… cut”). The fourth has a blue border and the Escolhida (Chosen) badge.
Four takes of the same line, and the one that made it into the reel.

At the end, the finishing skill delivers the project assembled in CapCut, with the recording, the scenes and the sounds on the timeline. I just open it, tweak what I want and export:

CapCut desktop with the project 10-abrindo-urna-pt01 open: in the media panel, the cut, the scenes and the sounds; in the player, the reel with the voting machine on top and the recording below; on the timeline, the video, scene and sound tracks already assembled.
The voting machine reel project, assembled by the skill, open in CapCut.

Install

The easiest way

You don’t need to do any of this by hand. Open Claude Code and ask: “install the eusouosalmo skills with npx skills@latest add eusouosalmo/skills and everything they need to run”. Or click “Copy for AI”, at the top of this article, and paste it into Claude Code: it follows the steps below on your computer.

What you need on your computer:

  • Node.js, to run the installer
  • ffmpeg
  • a Whisper with word timestamps (faster-whisper or whisper.cpp), to transcribe the recordings
  • Python with Playwright, to build the scenes
  • CapCut desktop, only if you want the reel as a CapCut project; the finished mp4 doesn’t need it

You don’t need to install everything up front. The first time, the animation skill checks what’s missing and shows the command for your system.

  1. In your computer’s terminal (on Windows, PowerShell), run:

    bash
    npx skills@latest add eusouosalmo/skills

    The installer asks which skills you want and for which agents. For reels, you don’t need skill-forge or git-guardrails, which are for engineering. Use the space bar to mark each one and Enter to confirm:

    The installer in the terminal: it clones the eusouosalmo/skills repository and shows the list of skills to mark, with Select All at the top.
    The installer lists the repository's skills for you to pick.

  2. In Claude Code, run /setup-eusouosalmo-skills once. It finds your folders (reels, scripts, style, CapCut) and saves the paths, so no skill asks again.

  3. Ask in plain language, for example: “cut the raw clips in this reel’s folder and drop the bad takes”.

From script to finished reel

This is the path I follow for every reel:

  1. Write the script. Ask the script skill, starting from your notes. It delivers the teleprompter text already timed and, for each block, what shows on screen:

    A script open in Obsidian, in a three-column table: the block with its word count, the line and the visual. The hook and the opening of the voting machine reel describe the machine in an exploded view and the 1996 machine with a floppy disk.
    The start of the voting machine reel script: the line and what shows on screen in each block.

  2. Record in order. Record vertically, following the order of the script, because the recording order becomes the video order. Repeat each sentence until it comes out the way you want. Messed up? Say “cut” and start the sentence again. The cutting skill separates the takes from the backstage talk, so you can talk to whoever is filming between takes.

  3. Put everything in one folder. One folder per reel, inside your reels folder, with the raw recordings and the script. I keep the raw clips in a brutos subfolder, just as they come off the phone:

    Windows File Explorer in the folder D:, Vídeos, 10-abrindo-urna, brutos, with 28 MOV videos from the phone as thumbnails, from IMG_1605 to IMG_1632.
    The 28 raw clips of the voting machine reel, in a subfolder of the reel's folder.

  4. Using CapCut? Create an empty project in it and close the app before asking for the edit.

  5. Ask for the full edit:

    Full edit

    Edit the reel in the folder [folder name] end to end. The script is in the folder. I want the CapCut project.

  6. Approve at the stops. The agent stops at three points: the takes table, the scene plan and the scene preview. Nothing is rendered before you answer.

  7. Open, tweak and export. Then ask for the post caption: “write the post caption for this reel”.

Note

If the session drops halfway, no problem: the edit logs its progress in an edicao.md file inside the reel’s folder, and a new agent picks up where it left off.

Make the workflow your own

The workflow comes with my way of speaking and editing. The idea is for you to swap mine for yours, on three fronts.

Your voice. The eusouosalmo-voice skill is the way I speak and write. Use it as a model to create a skill with your voice, otherwise your script comes out sounding like me. Gather a few texts of yours, like captions you’ve already posted or the transcript of a video, and ask:

Your voice

Use the eusouosalmo-voice skill as a model and create a skill with my voice from these texts of mine. Keep the same structure and replace the examples, the expressions and the rules with mine.

[paste three to five of your texts here]

Your visual identity. The scenes follow the colors and fonts of a design system, and the final assembly copies the editing style of one of your CapCut projects: the intro, the sounds, the zoom and the captions. Tell it which ones are yours:

Your visual identity

These are my brand’s colors and fonts: [colors in hex and font names]. Use these colors and fonts in the scenes of my reels. And use my CapCut project called [project name] as the editing template: intro, sounds, zoom and captions.

Your taste. When the animation skill delivers the scenes, it asks what you thought. Every reason you give becomes a rule in a style file, estilo.md, and it holds for every video after that. One of mine: I asked to see the frames of each scene before generating the video, because fixing a still image spends far fewer tokens than redoing the animation. Today the skill always does that. This was the scene preview for the voting machine reel, approved before the render:

Ten still frames of the voting machine reel scenes, side by side in two rows: the open machine, the timeline of the voting steps, the polling station report, the digital signature, the trip without internet, the tally and the part 2 card.
The scene preview: still images I approve before they become video.

Other examples of what you can say:

  • “Never cover my face with animation.”
  • “On-screen words come in lowercase.”
  • “My logo shows up only in the last second.”

Your file starts empty and becomes more yours with every video.

Today

Today I no longer explain the process for every video. I ask, approve what the skill shows me, and every correction is saved for the next one. This is the difference in one of my reels, at the same instant, before and after the workflow:

On the left, me talking to the camera in the raw recording. On the right, the same frame in the finished reel: on top, the animation of an electronic voting machine with a timeline of the voting steps; below, the recording, with the caption “ISSO OCORRA” between the two halves.
The raw recording and the finished reel, at the same instant.

All the skills are in the eusouosalmo/skills repository.