More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Podcast to Shorts: vertical clips with synced captions

    1

    Cut a long recording into checked vertical clips with word-by-word captions for Shorts, Reels, TikTok and LinkedIn.

    $12

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Podcast to Shorts: vertical clips with synced captions

    Podcast to Shorts: vertical clips with synced captions

    Example session with this skill installed

    Clip this podcast episode into vertical shorts with captions: episode-12.mp4

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

    Two vertical clips are ready to watch in shorts/clips/final/. Both passed the file checks and the listen-back check, and nothing has been posted.

    ClipSource timeLengthLayoutHook lineOpens onEnds on
    c010:06.6 to 0:35.128.6 sfit (whole frame)The podium test for life"We call this the podium test.""...above a background of abiotic chemistry."
    c021:57.7 to 2:39.441.6 sfull, centred cropThe hypothesis of last resort"Carl Sagan once said, life is the hypothesis of last resort""...that will be quite a challenge."
    Checkc01c02
    Size1080x19201080x1920
    Loudness (target -14 LUFS)-14.4-14.4
    True peak (limit -1 dBTP)-1.3-1.2
    Planned words heard in the render100%100%
    Caption sync, median0.02 s0.025 s
    Quiet after the last word0.26 s0.17 s
    ResultPASSPASS, 1 WARN

    What to know

    • c02 has one cut. It removes the doubled "the" in "any of the the possible", and the re-check heard that line once.
    • The c02 warning says "One" is 0.86 s off. I measured the audio there: the caption lights up as the word starts, and the second transcription had stretched the word back over the pause.
    • c01 shows the whole frame because the shot jumps between the speaker and b-roll, and in one shot he sits left of centre.
    • Your source file has the same SHA-256 before and after.

    Next steps

    • Watch both clips with sound. A PASS covers the technical checks only.
    • Before posting c02, confirm the Sagan quote and the rights to the archival footage it opens over.
    • Tell me which clips to keep, change or drop, and I will save any rule you agree to in editing-profile.md.

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

    About this skill

    Turning a one-hour episode into short clips usually means scrubbing a timeline for an afternoon. Clips from a rushed job tend to start halfway through a thought or lose the last word.

    This skill gives your agent a tested workflow for most of that job. It runs from the terminal with Python scripts and ffmpeg. The agent transcribes the recording with word timestamps, corrects each word's start and end against the actual audio, and proposes moments that make sense to a stranger. You approve the picks. Each clip renders at 1080x1920 with captions that light up word by word and an optional hook line, in one of three layouts: one speaker, two speakers stacked, or a screen demo shown whole. The crop is a fixed frame you set per clip. There is no face tracking, so a speaker who moves a lot needs a wider frame.

    Every clip is checked. A boundary check blocks any start, end or cut that lands inside a word. After rendering, the clip is transcribed again and compared with the plan: first and last words, caption timing against the voice, room after the last word, and leftover ums. Loudness is set to -14 LUFS with true peak at or under -1 dBTP. Your original file is never written. Rules you agree to are kept in an editing profile for the next episode.

    What you get: SKILL.md, four reference guides, five Python scripts, two templates and 20 tests on synthetic footage.

    Needs ffmpeg and Python with Pillow and numpy. Transcribe locally with faster-whisper or openai-whisper (free and private), or with any OpenAI-compatible Whisper endpoint and your own key. The network is used only for the model download on first run, the optional free font, or audio sent to the endpoint you pick.

    Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Video Skills.

    There is no face tracking: each clip uses one fixed crop, so a speaker who moves a lot needs a wider frame or the fit layout. It picks and checks clips but does not post them. You need ffmpeg and a transcription engine, either local whisper or your own Whisper endpoint. A PASS covers the technical checks only, so watch every clip with sound before posting.

    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 buying a skill.

    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 13 days ago

    Software
    FFmpegffprobePython 3.9+PillowNumPy
    faster-whisper
    oropenai-whisper
    Optional
    OpenAI-compatible Whisper API

    Local transcription is free. The Whisper API can be OpenAI, Groq or self-hosted; put its key in an environment variable.

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables
    Listed13 days ago

    What's inside

    Frequently Asked Questions

    ffmpeg and ffprobe, Python 3.9 or newer with Pillow and numpy, and one way to transcribe: faster-whisper or openai-whisper on your machine (free), or an OpenAI-compatible Whisper endpoint with your key in an environment variable. Captions are drawn with Pillow, so an ffmpeg build without libass or drawtext still works.

    Your agent reads the whole transcript and proposes moments against written rules: the opening makes sense cold, the payoff is inside, the clip ends when the thought ends. Each pick goes into clips.json with a score and a reason. You approve or reject every clip, and nothing counts as final until you have watched it with sound.

    No. Each clip uses one fixed crop for its whole length. When the speaker moves or the shot changes, use a wider crop box, split the clip, or the fit layout, which shows the whole 16:9 picture over a blurred copy of itself. A contact sheet of four frames per clip helps you spot a bad crop.

    With a local engine, only the one-time model download touches the network. With a hosted endpoint, the recording's audio is uploaded to the endpoint you pick, with the key you set. The optional font script downloads Inter Tight from the Google Fonts repository. Your original recording is never written to.

    The skill ships 20 tests on synthetic footage, which passed in about 14 seconds. For this demo it clipped a 3.5-minute public-domain NASA video into two shorts through Groq. Both passed size, loudness (-14.4 LUFS), true peak and the listen-back check. The run also found two problems in the transcription step with Groq, and both are fixed in the shipped version.

    A plain prompt usually gets you timestamps and an ffmpeg command. This adds scripts that move each word edge onto the real audio, block any cut that lands inside a word, light captions word by word, set loudness to -14 LUFS, then transcribe the finished clip again to check the first and last words, caption timing and leftover ums.

    New in Image & Video