FervorCreative AI
Live Latest 21.11.26 · morning 422 tools tracked 1040 workflows indexed 651 topics Hot: Claude Code, Gaussian splatting, ComfyUI

Video Studio's 40 free Claude skills make a styled reel from a raw clip, but the edit is decided in a readable plan file between transcription and render, and the creator who rewrites that plan gets a reel worth posting while the one who skips it gets a template.

Claude Video Editing SkillsVideo Studio CoreClaude CodeFFmpegai-editingcreative-workflowsvideo-genlocal-creative-ai

Turn a Raw Talking-Head Clip Into a Finished Reel by Naming a Look: 40 Free Claude Video Editing Skills

Video Studio is a free, open-source pack of Claude skills that cuts pauses, adds word-by-word captions, motion graphics, sound and a grade on your own computer. The step that makes the edit good is one most people will skip.

Most short-form talking-head edits are built from the same dozen moves. Tight jump cuts. Captions that light up word by word. A zoom punch on the word that matters. A hook box in the first two seconds. A giant headline that sits behind the speaker's head. Editors who make these for a living know the moves so well they could name them, and that is exactly what a new open-source project does.

Video Studio, released on GitHub on November 17 under the MIT license, packages 40 of these looks as Claude skills. Each one has a name, like halo-cut, depth-type, news-desk or vhs-tape. You put a clip in a folder, open Claude there, and say something like "Edit clip.mp4 with halo-cut. Cut every pause." Claude transcribes the clip, cuts it, drafts the edit, checks still frames, renders and finishes. You get final.mp4.

That sounds like a template app. It is not, and the difference is the most useful thing about it.

The edit lives in a file you can read

Template apps hide their decisions. You press a button and the captions land where the app decided. Video Studio writes its decisions down first.

Here is the order it works in, taken from the project's own documentation:

  1. Prep. The clip is converted to a steady 30 frames per second and turned upright.
  2. Transcribe. A free, local speech-to-text tool called faster-whisper writes down every word with its exact timing.
  3. Auto-cut. Pauses longer than about a third of a second, filler words and any retakes you name are removed.
  4. Cut out the speaker. For looks with text behind the head, a small model separates the person from the background.
  5. Draft the plan. A script writes plan.json, a plain text file listing every caption, key word, zoom, card and sound effect with its timing.
  6. Render. Each overlay is drawn as a web page, frame by frame, in a hidden browser, then layered onto your footage with FFmpeg, the free command-line video tool.
  7. Finish. Color grade, grain, sound effects, music that ducks under the voice, and loudness set to -14 LUFS, a common loudness target for streaming platforms.

Step five is the one that matters. The project's own core skill calls the automatic plan "a heuristic first draft" and puts it bluntly: refining the plan "is where the edit gets good."

Why the plan is where you earn your reel

The automatic draft picks key words by rules. Rules do not know which word in your sentence carries the point. They do not know that your hook should promise the payoff, or that the number you mention halfway through is the reason anyone will share the clip.

The core skill gives Claude a checklist for improving the plan before anything renders, and you can hold Claude to it or edit the file yourself:

  • The hook. Rewrite the first two seconds into seven words or fewer.
  • Key words. Replace weak automatic picks with numbers, results, names and contrasts.
  • One idea per element. Behind-the-head headlines get three words at most.
  • Numbers become graphics. Every important number turns into a counter or a chart.
  • Lists appear as they are spoken. Each item lands the moment the speaker says it.
  • Timing. Text appears a fraction of a second before the word and never stays on screen for less than 0.6 seconds.
  • Safe zones. For vertical video, text stays between 10% and 85% of the frame height and away from the buttons on the right edge.
  • Density. Something changes every 1.5 to 3 seconds in a reel, every 4 to 8 seconds in long-form, with no more than two overlays at once.

Read that list again as an editor. It is a decent short-form style guide. The value of Video Studio is less that it renders captions and more that it makes those editorial calls visible and editable as text before you spend time rendering.

Put this into practice: your first reel

Give yourself an hour for setup and the first edit. You need Claude Code or Cowork with skills enabled, Python 3.10 or later, and FFmpeg installed.

1. Install the tools once. Run the install command from the README for faster-whisper and the other Python libraries, then install the hidden browser it uses for graphics. The first transcription downloads a speech model of about 460 MB.

2. Install the skills. Clone the repository and run its installer, which copies the skills into your Claude skills folder. Claude Code users can add it as a plugin instead. Whichever route you take, the video-studio-core folder must come along, because every look depends on it.

3. Check your setup. Run the included check_env.py script. If anything is missing, it prints the exact fix.

4. Start with halo-cut. It is the cleanest of the talking-head looks: two or three uppercase words at a time, the spoken word in yellow, key words in cyan, a zoom punch on strong words, no more than one every 2.4 seconds. It is a good baseline before you try anything louder.

5. Stop at the plan. Ask Claude to run the edit only as far as the plan. Then open plan.json next to the transcript in words.json and rewrite the hook yourself. Pick the four to eight words that carry your message. Fix any misheard words, keeping their timings.

6. Look at stills before rendering. Ask for a contact sheet of a few frames. Check that text is readable, nothing is cut off and nothing covers the face unless you meant it to. This takes seconds and saves a full render.

7. Render and finish. Run the full render, then the finish step. Ask for extra sizes like 1:1 or 16:9 if you post to more than one platform. Add music by naming a track in the plan.

Once you trust the pipeline, try a look that changes the format, not just the captions. split-proof puts the speaker in the bottom half and a new card, number or clip in the top half for every claim. clip-miner finds the best 20 to 60 second moments in a long podcast and turns each into a captioned vertical short. pod-frame switches cameras by who is talking.

Make the look your own

Forty looks will feel generic fast if everyone uses them unchanged. The project anticipates this. Each look is a folder with two small style files, theme.json and theme.css. Copy a folder, rename it, change the colors, fonts and spacing, and you have a house style that Claude will apply the same way every time.

For a creator with a recognizable visual identity, this is the real payoff. Build one look that matches your brand, then every clip goes through the same pipeline with the same hook discipline, and the only variable left is what you said on camera.

Where it breaks

Here is what to expect before you rely on it.

It is two days old. At the time of writing the repository had 21 stars on GitHub. Forty looks means forty places for a rendering bug, and you will find some of them.

Rendering by web page is slow on long clips. Drawing every overlay frame by frame in a hidden browser is flexible but not fast. The core skill's own advice for slowness is a smaller transcription model and more parallel render workers on a strong machine. Expect that to be fine for a short reel and slow for a long video.

The speaker cut-out can wobble. Looks that put text behind your head depend on a person mask. The troubleshooting notes cover jumpy masks, harsh hair edges and the wrong person being picked, which tells you those things happen. A plain background and good light make it far more reliable.

Fonts need the internet. The looks pull fonts from Google Fonts. Offline, you get fallback fonts and a different look.

It needs a clip worth editing. Every look assumes a person talking to camera with something to say. It will not rescue a ramble. Tight jump cuts on a weak take produce a fast weak take.

Rights are on you. The project says it plainly: only use music, footage and reviews you have the rights to, and its testimonial look is for real customer quotes only. Do not fill it with invented praise.

The skill is still yours

Tools like this make the mechanical parts of a short-form edit close to free: the cutting, the caption timing, the loudness. What they cannot do is decide which seven words open your video, or which number in the middle deserves a chart.

Video Studio is honest about that, which is why it stops for a plan. Use that stop. Run one clip through halo-cut this week, refuse to render until you have rewritten the hook yourself, and compare the result with the automatic draft. The difference between the two is the part of editing that is still your job.


From the briefing