Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Transcript to Shorts

    1

    Turn a long video, podcast or livestream into the few clips that actually work as Shorts, TikToks and Reels.

    $15

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Transcript to Shorts

    Transcript to Shorts

    Example session with this skill installed

    Find the best 10-60 second clips in podcast-episode.srt for TikTok. The video is podcast-episode.mp4. Add burned-in captions.

    • Read your context and instructions
    • Compiled the transcript to shorts

    Summary

    I indexed podcast-episode.srt (32 segments, 2.2 minutes, 1 window) and read all of it. I proposed 5 moments, and the script kept 3. One candidate was rejected because its quote isn't in the transcript, and one was dropped because it overlaps a stronger clip. I used your 10-60 s range, with 9:16 framing and captions.

    3 clips (5 proposed, 1 rejected, 1 duplicate)

    #ScoreStartLengthTitle
    1880:00:17.349sThe $40 plugin that cost $9,000
    2860:01:47.823sRaise prices when sales are slow?
    3490:01:09.930sWhy I call every customer who asks for a refund

    1. The $40 plugin that cost $9,000

    • Why: Theo switches to a cheaper payment plugin, checkout silently fails for six hours, and he shares the one rule that would have caught it.
    • Risk: Runs close to 50 seconds; the test-order rule needs the last three lines to land.
    • Scores: hook 3/4, payoff 4/4, clarity 4/4, novelty 3/4, emotion 3/4, value 4/4
    ffmpeg -hide_banner -ss 17.390 -i podcast-episode.mp4 -t 48.820 -map 0:v:0 -map 0:a:0? -vf "scale=1080:1920:force_original_aspect_ratio=increase,crop=1080:1920,setsar=1,subtitles=clips/01-the-40-plugin-that-cost-9-000.srt:force_style='FontName=Arial,FontSize=13,Bold=1,Outline=2,Alignment=2,MarginV=60'" -c:v libx264 -preset veryfast -crf 20 -pix_fmt yuv420p -c:a aac -b:a 160k -movflags +faststart clips/01-the-40-plugin-that-cost-9-000.mp4
    

    2. Raise prices when sales are slow?

    • Why: A counter-intuitive answer backed by a real result: $12 to $19 and it sold more.
    • Risk: One anecdote; the advice to test first matters and must stay in.
    • Scores: hook 4/4, payoff 3/4, clarity 4/4, novelty 4/4, emotion 1/4, value 4/4

    3. Why I call every customer who asks for a refund (score capped at 49)

    • Why: Calling refund requests reveals confusing instructions; rewriting one page cut refunds by a third.
    • Risk: The hook arrives about 7 seconds in, so the rubric caps it.

    Fix: start at segment 18 ("It was, at first. But here's what surprised me.") and it becomes a strong clip.

    Not used

    • Intro: the quote isn't in the selected segments (rejected)
    • Zero orders on a Friday night: overlaps "The $40 plugin that cost $9,000"

    Next steps

    1. Say "cut them" and I'll run the three ffmpeg commands. They write to clips/ and never touch the source.
    2. Or I can move clip 3 to start at segment 18 and re-score it first.
    3. If the speaker isn't centred, give me the crop offset (0.0 left to 1.0 right) and I'll adjust the commands.

    Scores are editorial judgements from the transcript, not view predictions.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Turn a long video, podcast or livestream into the few clips that actually work as Shorts, TikToks and Reels.

    Give your agent a stream link (YouTube, Twitch or Kick) or a transcript (SRT, VTT or Whisper JSON). It reads the whole thing and proposes moments. Then a bundled script checks every clip against the transcript before you see it:

    • Starts from a link: paste a URL and the agent fetches only the captions with yt-dlp, so you don't download hours of video to find the clips. It downloads just the ranges you decide to cut.
    • Proven clips only: each clip needs a quote found word-for-word in the selected lines, so invented or misplaced moments are rejected.
    • Editorial rubric: hook, payoff, clarity, novelty, emotion and value, each tied to a specific line. Scores the lines don't support are capped automatically. A hook that arrives 8 seconds in can't rank as strong.
    • No duplicates: overlapping clips and the same story told twice are dropped.
    • Long recordings covered: multi-hour transcripts are split into overlapping windows, and the agent reads every one. Messy YouTube auto-captions (repeated rolling lines) are cleaned automatically.
    • Cut-ready: ffmpeg commands for 9:16, 1:1, 4:5 or 16:9, plus per-clip caption files to burn in. Nothing runs until you say so.

    What's in the zip

    • SKILL.md: the step-by-step workflow your agent follows
    • scripts/segments.py: parses SRT, VTT or Whisper JSON and splits it into overlapping windows
    • scripts/clips.py: validates, scores, deduplicates, writes caption files, prints ffmpeg commands
    • reference/: the scoring rubric, the candidate format, and ffmpeg export recipes (off-centre crops, letterboxing, caption styling, loudness)
    • examples/: a sample transcript, candidates and the real report they produce

    Limits

    Scores judge the words, not the picture. The skill can't see faces or on-screen text, or hear laughter, so check each clip before posting. Scores are editorial judgements, not view predictions.

    Built from the clipping engine of Cutroom (cutroom.batuukka.com), an open-source clipping studio used on multi-hour livestreams. The scripts are pure Python standard library and make no network calls. Starting from a link needs yt-dlp. Clip only content you're allowed to use: your own, or a creator's with permission.

    Needs captions or a transcript. From a link, the video must have captions (auto-generated ones work); otherwise provide SRT, VTT or Whisper JSON. No audio or visual analysis: scoring is based on the text only. ffmpeg must be installed to cut clips, and yt-dlp to start from a link.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 7 days ago

    Requires

    Python 3.8+
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access

    File Scopes

    clips/**

    Runs two local Python scripts (standard library only). Reads the transcript you name; writes segments.json, candidates.json and, with --captions, one caption .srt per clip in the output folder. Prints ffmpeg commands but never runs them without your OK. Network: only when you give a link. Your agent then runs yt-dlp to fetch that video's captions, and the video ranges you ask it to cut. No environment variables.

    Listed8 days ago

    What's inside

    Frequently Asked Questions

    Either. Paste a YouTube, Twitch or Kick link and your agent fetches the captions with yt-dlp. The video needs captions, and auto-generated ones work. Or provide a transcript with timestamps in .srt, .vtt or Whisper JSON. The skill works from the text and timing, so a cleaner transcript gives better clips.

    Your agent scores each candidate 0-4 on hook, payoff, clarity, novelty, emotion and value, and must point to the exact line behind each score. The script then caps any score those lines don't support: a hook more than 5 seconds in, a long tail after the payoff, emotion without an earlier contrast. It rejects clips whose quote isn't in the transcript. These are editorial judgements, not view predictions.

    Both. It prints a ready ffmpeg command per clip that cuts the moment, crops or pads it to 9:16, 1:1, 4:5 or 16:9 and can burn in captions. Your agent runs them only when you say so, writing new files to clips/ and never touching the source video.

    SKILL.md (the workflow your agent follows), two Python scripts (segments.py to parse and window the transcript, clips.py to validate, score, deduplicate and generate commands and captions), reference docs for the rubric, the candidate format and ffmpeg recipes, and a worked example with a transcript, candidates and the real report they produce.

    No, the skill relies entirely on text and timing data. It identifies high-impact moments based on narrative structure and dialogue, making it perfect for podcasts, interviews, and livestreams where the message drives the content.

    Yes. segments.py splits long transcripts into overlapping windows of about 900 words, so moments near a boundary keep their context, and the agent is told to read every window rather than sample. Repeated lines from YouTube auto-captions are removed first. Overlapping or near-duplicate clips from different windows are then dropped by clips.py.

    New in Image & Video