Repurpose a Long YouTube Video Into Shorts With Claude

To repurpose a long YouTube video into Shorts with Claude, you do not need a video editor with a timeline scrubber open all afternoon. You need the transcript, a way to score candidate moments, and a publish path that puts each Short on the right platform with the right framing. Claude can do the reading, ranking and writing. Creator OS handles the cutting, the vertical reframe, and the upload. This post walks through one workflow, start to finish, for a single long video.

  long video (mp4 / YouTube URL)
        |
        v
  transcript + chapters -------> Claude ranks moments
        |                              |
        |                              v
        |                     shortlist with scores
        |                              |
        v                              v
  OPEN VIDEO EDIT skills  <---- cut + reframe vertical
        |                              |
        v                              v
  mp4 per Short -----------> Creator OS upload + schedule
                                       |
                                       v
                        Instagram Reels / TikTok / Shorts

Why Claude Beats Manual Shorts Repurposing

Watching a 40 minute video back at 2x to find the good 45 seconds is the expensive part. It is also the part a language model is genuinely good at. A transcript is text. Finding the moment where someone says something that lands is a reading task, and reading is what you are already paying Claude for.

The second reason is consistency. When you pick clips by feel, you pick the ones you remember, which means you pick from the first ten minutes. A model that reads the whole transcript picks from minute 34 too. That is usually where the best clip was hiding.

The third reason is the publish path. Creator OS ships a hosted MCP server and a CLI, so a Claude session can create the post, attach the media, schedule it, and read the analytics afterward without you switching tabs. Details are at the MCP docs.

What You Need Before You Start

Four things, and none of them are heavy:

• The long video. Either a local file or a YouTube URL you own.
• A transcript with rough timestamps. YouTube auto captions are fine to start.
• Claude with the Creator OS MCP server connected. In Claude, go to Settings, then Connectors, then Add custom connector, paste the URL, sign in, pick your workspace, and choose read only or read & write.
• The open source editing skills, which render locally. The repo is github.com/kevinbadi/open-edits.

If you would rather watch this run once before you try it, the walkthroughs live on the KevBuildsApps YouTube channel, and the launch video covers the open source tooling at this link.

One thing to get right up front: read only connections only see read tools, and the API refuses writes from them. If you want Claude to actually post, connect read & write.

Step 1: Extract the Transcript and Chapter Markers

Pull the transcript first. If the video has chapters in the description, grab those too. Chapters are a free hint about where the topic shifts, and topic shifts are where complete thoughts start.

Then clean it. Auto captions have no punctuation and mangle product names. Two minutes of cleanup saves you a worse problem later, because Claude will happily quote a sentence that makes no sense.

Give Claude the transcript as a file or paste it with timestamps every 15 to 30 seconds. Timestamps are what let you cut later without guessing. If your transcript has no timestamps, add them with a rough pass before you continue. A clip that starts four seconds late feels broken no matter how good the line is.

Step 2: Ask Claude to Find the Highest-Value Moments

Now prompt for candidate moments, not finished clips. Something like:

Read this transcript. Find every moment that could stand
alone as a 30 to 60 second video for someone who never
saw the full episode.

For each one give me:
- start and end timestamps
- the exact opening line
- why it stands alone
- what the viewer learns

Skip anything that needs earlier context to make sense.
Give me 15 candidates, ranked.

Two phrases in that prompt are doing real work. Stand alone is the constraint that kills most bad clips. Needs earlier context is the filter that removes them.

Expect Claude to return more candidates than you need. That is correct. You want 15 so you can throw away 10.

If your long video is a course or a talking head recording, the same transcript pass works. The workflow in book to course with a Claude skill uses the same idea: read the source, find the structure, then split it.

Step 3: Score and Shortlist Clips by Hook Strength

Fifteen candidates is not a decision. Score them. Ask Claude for four numbers per clip, each 1 to 5:

Sort by the total, keep the top 5 to 8. Then do the check a model will not do for you: read each opening line out loud. If you would scroll past it, so would everyone else. Cut it there and you have saved yourself a render.

This is also the moment to decide which clips are vertical by nature and which need a b-roll card behind them. A clip where the speaker is just talking usually needs something visual. A clip where the speaker is already showing a screen does not.

Step 4: Cut and Reframe Vertically From Claude

Cutting happens locally. OPEN VIDEO EDIT is the open source repo at github.com/kevinbadi/open-edits, and it covers the editing skills that turn a raw talking head clip into an animated short, with captions, logos, faces, GIFs, b-roll cards and counters. It renders with Python and ffmpeg on your own machine. The skills include split-animated-talking-head, longform-animated-talking-head, fireship-style-edit, fireship-style-short, split-animated-short and reverse-engineer.

So the loop is: Claude picks the moment from the transcript, you hand the timestamps to a skill, the skill cuts and reframes, and you get an mp4 per Short. The reframe is the part that matters. A 16:9 frame shoved into a 9:16 canvas with black bars is a scroll-past. Face tracking to a vertical crop is not.

Read more about the editing side in the OPEN VIDEO EDIT update.

One practical note: render one clip first and watch it on a phone before you render eight. A framing choice that looks fine on a monitor can crop someone’s hands off.

Step 5: Generate Captions, Titles, and Hooks

Each platform wants a different wrapper around the same clip. Instagram Reels captions have no clickable links, so links go in the first comment. TikTok allows up to 4,000 characters in a photo post description and needs consent flags set. YouTube Shorts takes a title up to 100 characters and a pinned first comment up to 10,000 characters. Threads caps a post at 500 characters.

Ask Claude for one pass per platform, and let Creator OS check the limits. The CLI has a length validator:

creatoros validate:post-length --text "your caption here"

And when you are writing from an MCP session, the tool for the same job is check_caption_length. Use it instead of counting characters by hand, because the number that matters is the platform’s, not yours.

For hooks, ask for five options per clip and pick one. Do not average them into a bland sentence. The hook is what the algorithm tests first, and a mediocre one wastes the edit.

If you are also cutting horizontal versions for a wider feed, the same source clips work. The open source Claude video editor walkthrough covers that path.

Step 6: Upload and Schedule Every Short From One Chat

Now the publish step, from the same conversation. Give Claude the path to each rendered file and tell it where each one goes. A prompt like this works:

For each of the 6 clips in ./shorts:
upload the file, then create a post
scheduled 3 days apart starting Monday 9am ET,
instagram as a Reel, tiktok, youtube as a Short.
Title from clip-title.txt, caption from
caption-instagram.txt for instagram.
Add the link to the first comment.
Show me the list before anything is scheduled.

Two tools do the heavy lifting here: upload_media_from_url and create_post. Uploading media once gives you a media id you can reuse on every platform, which is why you upload first and post second rather than attaching the same file six times.

If you prefer the terminal, the equivalent is a single command per clip:

creatoros posts:create --text "caption" \
  --platforms instagram,tiktok,youtube \
  --media med_abc123 \
  --title "Short title" \
  --scheduledAt 2026-03-09T13:00:00Z --timezone America/New_York

Networks you named that are not connected come back in missing_platforms, and the rest still go out. So a typo in a platform name does not silently drop the whole post.

Before you walk away, ask Claude to list what it scheduled. Then check your analytics a week later from the same chat. Per post analytics, daily account metrics, follower stats and best time to post all come back through the connection, so you can ask which hook style actually held people past three seconds. More on that in the YouTube Shorts automation post.

For cross platform variants of the same clip, the Reels to TikTok to Shorts guide goes deeper on what changes between them.

Common Mistakes That Kill Retention on Repurposed Shorts

Starting the cut at the sentence. Start two seconds earlier, on the intake of breath. The sentence is the payload; the setup is what keeps people from swiping.

No visual change in the first three seconds. A static face for three seconds reads as a still image. That is what the b-roll cards and counters in the editing skills are for.

Reusing the long video’s thumbnail framing. Shorts do not take custom thumbnails on YouTube. Design for the vertical crop instead of hoping the wide shot survives it.

One caption for every platform. Different lengths, different link behavior, different audience. Write once per platform, validate the length, then schedule.

Publishing all six on day one. You learn nothing and you burn the queue. Spread them out and read per post analytics between them.

Cutting so tight that the payoff is gone. A 20 second clip of a 60 second idea is not a Short, it is a fragment. Keep the ending.

Get started with Creator OS

The Creator plan is $19.99/month or $59.99/year and covers up to 8 connected accounts, which is 7 socials plus Skool. API keys, the MCP server, the CLI and the agent skills are included in every plan, so the workflow above does not need an upgrade.

Connect your accounts, add the MCP server, render your first Short with OPEN VIDEO EDIT, and schedule it from the same Claude chat. The full setup steps are in the docs, and you can start at creatoros.ca/sign-up.

If you want to see the whole thing run end to end before you set it up yourself, the KevBuildsApps YouTube channel has the walkthroughs.