More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIDeepSeekDeepSeekMuseMuse+22 more

    Remove Retakes, Stutters and Dead Air from Videos

    1

    Your agent finds the flubbed takes, keeps the last good one, lets you hear every join, then renders a clean copy.

    $9

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Remove Retakes, Stutters and Dead Air from Videos

    Remove Retakes, Stutters and Dead Air from Videos

    Example session with this skill installed

    Cut my retakes out of tutorial-04.mp4 and keep the last take of each line

    • Read your context and instructions
    • Compiled the remove retakes stutters
    • Generated the image

    Your retakes are out. tutorial-04 - tightened.mp4 runs 37.1 s, down from 56.7 s, and the original file is unchanged.

    #Removed (source time)What was cutPause left at the joinConfidence
    10:06.3 to 0:09.5False start "In column A, type the names of your"0.76 shigh
    20:14.7 to 0:20.8First take "Next, select column B and set the format to", then "sorry, let me redo that"0.76 shigh
    30:25.4 to 0:29.0First take "Now add a total at the bottom with the sum function."0.79 shigh
    40:33.6 to 0:37.4Two false starts of "To see where your money goes"0.62 slow (edge moved 0.17 s)
    50:43.07 to 0:43.13First "the" in "the the style"no pause to uselow
    60:49.1 to 0:52.0A 3.7 s pause after "save the file", cut down to 0.79 s0.79 slow
    CheckResult
    Source untouchedpass
    Length against the plan37.059 s against 37.026 s, pass
    Audio and picture end together0.026 s apart, same as the source
    Joins in quiet sound6 of 6 pass, join 5 listed for a listen (-33.2 dB)
    Re-transcript of the outputreads as one take, each line once

    Two things need your ear

    • Join 5 sits at 0:26.4 in the new file. Lining the cut up with the video frames shrank it to 0.067 s, so about 60 ms of the first "the" is still there. The re-transcript hears "Pick the style", but a listen settles it.
      (The current version rounds stutter cuts outward to whole frames, so this cut now removes the full span.)
    • Cut 6 shortens a pause you did not mention. If something was happening on screen during it, untick it.

    The review page is tutorial-04 - cut review.html. It needs the local server, so ask me to run build_review.py again when you want it open.

    Next steps

    • Press "Play every join" on the review page and tell me what to change.
    • If join 5 sounds clipped, I can widen that cut by a frame on each side and render again.
    • Once it sounds right, bring the tightened file into your editor for zooms, captions and music.

    review.html

    PNG · 1536×1024

    Generated

    Example file from a real run - the skill writes it into your workspace.

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

    About this skill

    If you record tutorials or talking-head videos in one pass, you flub a line, pause, say it again, and later spend an hour scrubbing the timeline for your mistakes. This skill hands that job to your agent.

    It transcribes the recording with word timestamps, finds the places where you restarted a line, and cuts from your first attempt to the last good take. It also catches doubled words and can shorten long silences. Before anything renders, a review page in your browser lets you hear each join, untick a cut, and adjust the times.

    Whisper timestamps are often off by a few tenths of a second, which is how automated cuts clip the last word of a sentence. Every cut edge is moved into silence measured from the audio itself, and each join is sized to your own pause rhythm. The render uses hard picture cuts with tiny audio fades, so there are no clicks, and frame-locked edges so sync holds across dozens of cuts. A final check confirms quiet joins, the planned length, and audio and picture ending together.

    What you get: instructions for your agent, eight Python scripts, the review page, two reference guides, and a self-test. It was tested on a real talking-head recording and on generated speech.

    Needs ffmpeg, Python 3.9+, and either local faster-whisper or whisper (free, private) or a key for an OpenAI-compatible Whisper API such as OpenAI or Groq. It runs from the terminal through your agent, with Python and ffmpeg; there is no app window beyond the review page. Network use: a local model downloads once on first run, an API engine uploads the audio track to the endpoint you choose, and the review page is served on localhost. Your source file is never changed.

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

    It finds retakes from repeated words, restart cues and pauses, so a line said again in completely different words may be missed. It never judges what is on screen. A stutter with no pause is cut at its quietest point, so that join is flagged for a listen. It makes hard cuts only: no zooms, captions or music.

    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

    Accounts
    Local Whisper
    orWhisper API key
    Software
    FFmpegffprobePython 3.9+

    Run faster-whisper or openai-whisper locally (free, offline), or set an OpenAI-compatible Whisper API key in an environment variable such as OPENAI_API_KEY or GROQ_API_KEY.

    Permissions

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

    What's inside

    Frequently Asked Questions

    ffmpeg and ffprobe, Python 3.9 or newer, and one transcription engine: faster-whisper or openai-whisper on your machine (free, private after a one-time model download), or an OpenAI-compatible Whisper API such as OpenAI or Groq, with the requests package and your key in an environment variable. numpy is optional and only speeds up level measurement.

    When a line is delivered more than once, the last delivery is the keeper, and the cut runs from the start of the first attempt to the start of the keeper. A script lists repeated phrases, doubled words, restart cues and long silences as leads. Your agent still reads the whole transcript, and doubtful spans are marked low confidence or left in.

    Each cut edge is moved into silence measured from the audio, never across a word, and the pause left at each join is sized to your usual pause. A stutter with no gap is cut at its quietest point and flagged for a listen. When cuts are lined up with the video frames they are rounded outward, so a stutter cut never shrinks and leaves no piece of the doubled word behind.

    Yes. A local review page plays the moment before each cut and the moment after it, lets you untick a cut or edit its times, and copies the edited list back as JSON for your agent. It is served on 127.0.0.1 so the video can seek. After the render, a check confirms quiet joins, the planned length, and audio and picture ending together.

    With faster-whisper or openai-whisper, only the one-time model download uses the network. With an API engine, the audio track is uploaded to the endpoint you choose. The review page is served on localhost and loads nothing from outside. Your source file is never changed, and every output gets a new name.

    The skill's self-test passes all 18 checks, including a replay of the demo's stutter cut. The demo used 57 seconds of generated speech with four planted retakes, a doubled word and a long pause, transcribed through Groq. Every planted problem was found, the video went from 56.7 to 37.1 seconds, the render check passed, and the new transcript read as one clean take.

    New in Image & Video