Inizia
Clippa - Turn Any Stream into Upload-Ready Shorts

Clippa - Turn Any Stream into Upload-Ready Shorts

Finding the moments is half the job. Cutting, reframing, captioning, titling and tagging each one is the other half. Clippa does both on your own computer, for whatever streams or posts you want to clip
#Video#Social media#Gaming
Valutazione
Servono altre valutazioni
Venduti
0
Come si usa
Scarica

You choose the topic
Tell Clippa what your content is about: a game (Valorant, Fortnite, Minecraft, GTA 6, Call of Duty, Roblox, League, Apex, Counter-Strike), or a niche (IRL / just chatting, podcast, comedy, fitness, finance, tech), or anything of your own ("my cooking stream"). The topic sets the hashtags written for every clip and gives moments that mention your topic a ranking boost. Don't like the hashtags? Pass your own and they replace the defaults. A clip can also carry its own tags.

What it does for every clip

  1. Picks the strongest moments from a timestamped transcript (hook, energy, pacing, clean start and end), or uses cut points you give it. No transcript? Turn on the optional built-in transcription and it works from the video alone.
  2. Cuts the clip and reframes it to 1080x1920 (9:16): the full frame over a blurred fill (best for 16:9 gameplay and streams) or a cropped fill with a left-to-right focus control.
  3. Burns in a short hook for the first ~3 seconds and large outlined captions: word-timed with the optional transcription, line-timed from your own transcript, or written by you.
  4. Normalises loudness toward -14 LUFS (single pass, so it can land a couple of LU off) and exports H.264 + AAC at 30 fps with faststart, the format the short-form apps accept.
  5. Writes a text file next to each clip with title, description, hashtags and the source time.

What is included

  • SKILL.md: a step-by-step workflow for your AI agent (ask the topic, check ffmpeg, get a transcript, pick clips, write clips.json, preview with a dry run, cut, verify with ffprobe and a frame check, report).
  • make_clips.py: the cutter (Python standard library plus ffmpeg).
  • clip_finder.py: the moment ranker. topics.py: topic to hashtags and keywords (15 presets, any custom topic).
  • transcribe.py: OPTIONAL local speech-to-text with word timestamps (uses the open-source faster-whisper package, installed by you only if you want it).
  • Setup guide, clips.json format, troubleshooting, two sample transcripts, an example clips.json, test cases.
  • Guardrails reference: no stating rumours, leaks or unverified claims as fact; health and money caveats; copyright and monetisation cautions.

Requirements

  • Python 3.8 or newer.
  • ffmpeg and ffprobe installed and on your PATH (the full build with libass). ffmpeg is not included; the setup guide shows how to install it.
  • Your own video file, plus either a timestamped transcript (SRT, VTT, or "12:30 text" lines), timestamps of the moments you chose, or the optional transcription below.
  • OPTIONAL transcription: pip install faster-whisper. The first run of each Whisper model downloads it once (about 75 MB tiny, 150 MB base, 500 MB small, 3 GB large-v3), then it runs offline. Your audio stays on your computer. It uses a CUDA GPU if the libraries are installed and falls back to CPU otherwise (slow on hours of video; fine for short clips).
  • Tested on Windows 11 with Python 3.12, ffmpeg 9.0 (full build) and faster-whisper 1.2.1. Other systems should work but are untested.

Good to know

  • Tested on a real full-length gaming video: clips came out as 1080x1920 H.264/AAC with captions, hooks and hashtags, checked with ffprobe and by eye.
  • It cannot download or watch videos. You supply the video file. Transcription is optional and off by default; nothing is installed or downloaded unless you choose it.
  • Topic presets are convenient defaults, not researched trend data, and hashtags do not guarantee reach. Edit or override them freely.
  • One continuous cut per clip (no jump cuts or internal trims); pick tighter in and out points or trim in an editor.
  • Captions from your own transcript are timed per line and can drift on fast talking. Word-timed captions from the optional transcription are only as accurate as speech recognition: expect mistakes on names, slang and loud game audio, so check them before posting.
  • No subject tracking: check the crop focus on action-heavy footage.
  • The moment picker reads text only, so it cannot see the footage: it favours spoken lines and reactions, and on story or cutscene playthroughs it tends to pick dialogue. Review the picks before posting.
  • If your video already has burned-in subtitles or a facecam, they stay in the frame and Clippa's captions are added on top; use --no-captions or crop mode for those.
  • Speech recognition can mishear names and slang; always check captions.
  • Nothing is uploaded or posted for you, and your source video is never modified.
  • Nobody can guarantee views. This removes common reasons clips fail; it does not promise results.
  • You are responsible for having the right to clip and post the footage you process.
  • Game and brand names appear only as example topics. Not affiliated with, endorsed by or sponsored by any game publisher, streamer or platform.

AGENT CARD - CAPABILITIES (accurate list)

  • Asks for the topic and maps it to hashtags and ranking keywords (presets or custom).
  • Ranks clip-worthy moments in a timestamped transcript, or in one it makes itself with optional local transcription (faster-whisper).
  • Cuts clips from a local video file into 1080x1920 H.264/AAC MP4s using ffmpeg (blur-fill or crop).
  • Burns in a hook and word-timed, line-timed or hand-written captions; normalises loudness.
  • Writes a title, description and hashtags file per clip; verifies output with ffprobe and a frame check.
    Does NOT: download or watch video, post or upload anything, install anything without the user's say-so, guarantee views, reach or income, state rumours or unverified claims as fact, invent leaks or quotes, modify the source file, or help evade copyright detection.

DATA SHARING DECLARATION

  • The skill sends no user data to any server and needs no credentials or API keys.
  • make_clips.py reads the user's video and transcript files and writes new files (mp4, ass, txt) only into the chosen output folder. It runs ffmpeg and ffprobe locally.
  • clip_finder.py and topics.py read local text or arguments and print a result; they write nothing.
  • By default: no downloads, no network access, no third-party Python packages. ffmpeg is installed separately by the user.
  • OPTIONAL transcription (--transcribe, transcribe.py), only when the user turns it on: it uses the third-party package faster-whisper, which the user installs with pip, and a Whisper model that is downloaded once from the Hugging Face Hub on first use. Audio is transcribed locally and is not uploaded anywhere. The skill never installs or downloads these on its own.