gta-clipper
Turns a finished GTA 6 stream (a Twitch or Kick VOD, a Twitch clip, a YouTube link or a file) that the user has permission to clip into vertical Shorts, with the facecam on top, the gameplay below, captions and zooms. It picks the best moments, shows a ranked list for approval, then cuts them. Use when the user runs /gta-clipper or asks to clip a GTA or gaming stream.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/itsme23476--gta-clipper)HTML
<a href="/listings/itsme23476--gta-clipper"><img src="/api/listings/itsme23476--gta-clipper/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
Paste this repo's link into Claude Code and say: "Install this skill and set it up for me."
Or do it by hand:
git clone https://github.com/Itsme23476/gta-clipper ~/.claude/skills/gta-clipper
The skill must live in `~/.claude/skills/gta-clipper`. Restart Claude Code after installing. It's separate from [Claude Clipper](https://github.com/Itsme23476/claude-clipper), so you can have both.
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 98 lines
- ---
- name: gta-clipper
- description: Turns a finished GTA 6 stream (a Twitch or Kick VOD, a Twitch clip, a YouTube link or a file) that the user has permission to clip into vertical Shorts, with the facecam on top, the gameplay below, captions and zooms. It picks the best moments, shows a ranked list for approval, then cuts them. Use when the user runs /gta-clipper or asks to clip a GTA or gaming stream.
- ---
- # GTA Clipper
- Everything runs on this computer except picking the moments, which is your own reading of the transcript.
- ## Rules
- - Only clip streams the user owns or has permission to use, such as a streamer's clipping campaign. If it looks like someone else's stream, ask first.
- - Never cut anything until the user approves the ranked list.
- - Ask before installing or downloading anything.
- ## Steps
- **1. Prepare.** Run this with the stream link or video file:
- ```bash
- python3 ~/.claude/skills/gta-clipper/clipper.py prep "<LINK OR VIDEO>"
- ```
- On Windows, use `python` wherever these steps say `python3`. On Windows and Linux, faces and captions need OpenCV and Pillow: if it says they're missing, ask before running `pip install opencv-python-headless pillow`.
- It takes a Twitch VOD (`twitch.tv/videos/<id>`), a Twitch clip, a Kick VOD (`kick.com/<channel>/videos/<id>`), a YouTube link or a file. For a link it downloads only the audio now, and each approved clip's video when it renders, so a 14-hour stream never fills the disk. A long stream takes a while to download and transcribe, so run it in the background and tell the user.
- If it says the stream is still live, tell the user to come back once it has ended. If yt-dlp says the VOD is for subscribers only, it needs the user's Twitch login: they can download the VOD themselves and give you the file.
- If it says no Whisper model was found, search for an existing `ggml-*.bin` file and pass it as `WHISPER_MODEL=<path>`. Only offer to download one if none exists. It is about 150 MB: `mkdir -p ~/.cache/whisper && curl -L -o ~/.cache/whisper/ggml-base.en.bin https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin`. Add `--out <folder>` to choose where it works. The default is `~/Desktop/Clips/<stream title>-<id>/`. Running it again on the same link reuses what it already downloaded.
- **2. Pick the moments.** Read all of `transcript.txt`. Each line starts with its time in seconds, and a `-` or `>>` marks a new speaker. A stream transcript is long, so read it in parts with the Read tool's offset and limit, and keep a shortlist as you go. If one line repeats over and over, Whisper lost that stretch under game music: skip it. Choose the best moments, 20 to 60 seconds each unless the user says otherwise:
- - The moments a viewer would clip: big reactions, chases, crashes, funny fails, run-ins with NPCs or the police, and wins. Skip dead air, menus, loading screens, reading chat, and ad or sponsor reads.
- - It makes sense to someone who never saw the stream, with no missing context.
- - It starts on a strong first line, never in the middle of a thought.
- - It ends on a payoff, not trailing off.
- - Start and end on sentence boundaries from the transcript.
- You only have what was said, so a moment that only looks good on screen won't show up. Score each out of 10, and be strict. A long stream might have ten good moments, not fifty.
- **3. Show the ranked list, ask for a style, and stop.** A table with rank, score, start time, length, title, and the opening line. In the same message, ask which style to use. Each file in `~/.claude/skills/gta-clipper/styles/` is one style:
- - `hormozi` (the default): bold white captions, punch words in giant capitals and key words in red script, and a title box for the first two seconds.
- - `elegant`: lower-case captions that build word by word in stepped lines, key words in gold serif, and a title box for the first two seconds.
- - `classic`: bold outlined capitals with yellow highlight words.
- - `podcast`: small white subtitles that build word by word, and any caption holding a key word jumps up in size, with punch words in heavy capitals and emphasis words in a bold serif. A black title box with a red word for the first two seconds. Its accent colours are red `#E44848`, gold `#F0C048` and blue `#82C8FA`.
- - Any other file there is a style the user taught it. Describe it from its keys.
- Wait for the user's OK on both. They may drop or change some clips.
- **4. Save the picks** to `picks.json` in the same folder, with the chosen style as `preset`:
- ```json
- {"preset": "hormozi",
- "fix": {"Leo Nida": "Leonida"},
- "clips": [{"rank": 1, "score": 9, "title": "Short title", "caption": "Post caption with hashtags",
- "start": 123.4, "end": 171.2, "game": [[125.0, 131.5], [140.2, 144.0]],
- "highlights": ["cops", "crashed"], "hook": "HE DID NOT SEE *THAT* COMING...",
- "emphasis": ["no way", "get out", {"wasted": "#F0484E"}]}]}
- ```
- `game` lists the stretches where the game's characters talk, as `[start, end]` in seconds: a line's start and the next line's start in the transcript. Those words get no captions, because GTA shows its own subtitles, so the captions are only the streamer. Cutscenes, NPCs and phone calls are game dialogue, and so is anything written like a script. Leave a line in when you can't tell.
- `highlights` is one or two words or short phrases the streamer says at the clip's payoff. The first time each is said, the facecam and the gameplay punch in together for a second and a half, at most twice a clip, so put the biggest one first. In `hormozi` they are also the punch words in giant capitals, so prefer one-word payoffs.
- `emphasis` is the key word or short phrase in most of the streamer's sentences, about one per caption. Highlights and emphasis must be the streamer's words, not the game's. Each one gets the accent colour and font, without a zoom: red script in `hormozi`, gold serif in `elegant`, a big white serif in `podcast`. Write `{"word": "#hex"}` to give a word its own colour. Phrases match only as a whole, so "get out of here" never colours every "of".
- `hook` is a teaser of up to six words in capitals, shown in a box for the first two seconds, such as `HE CRASHED *ON PURPOSE*...`. Make it a cliffhanger, not a summary, and wrap the word to colour in `*asterisks*`.
- Whisper mishears names, GTA 6 ones included: Leonida, Vice City, Lucia (often "Lucy"), Jason, and the streamer's own name. It writes "chat" as "Chad" when a streamer talks to chat. Read for that and add a top-level `"fix"`. Put it on a single clip instead if a word is only wrong there.
- **5. Render:**
- ```bash
- python3 ~/.claude/skills/gta-clipper/clipper.py render "<FOLDER>"
- ```
- For a link, the first render of each clip downloads just that clip's video. Clips land in `<FOLDER>/clips/`, with titles and captions in `post-copy.md`. Open at least one clip's frames to check the result before saying it's done.
- ## Changes
- Edit `picks.json` and re-render only what changed with `--only 2`. For one clip, set `"zoom": false`, change `start` or `end`, or add `"style": {"highlight": "#00AAFF"}`. To change the whole batch, switch `preset` or add a top-level `"style"` with the keys to change.
- The layout is chosen for you: `split` puts the streamer's facecam on top and the gameplay below, with the captions on the seam, for a stream with a facecam on the left or right side, in a corner or halfway up. `crop` follows a talking head, such as a Just Chatting stream, and `fit` is for gameplay with no facecam. Set `"layout"` on a clip to override it.
- Split keys: `face_height` is the facecam's share of the height (0.4, so the gameplay gets 60%). The top panel is the facecam's own box, found where the room stays still while the game moves, and centred on the face as far as the box allows. `cam_size` scales it (1): lower it, such as 0.9, if a strip of gameplay shows at an edge, or to zoom in on the face.
- Style keys: `font`, `size`, `color`, `highlight`, `stroke`, `stroke_width` (0 for no outline), `shadow` (blur in pixels, 0 for none), `y` (caption height, 0 to 1, except in `split`, where captions sit on the seam), `max_words` (words per caption), `upper`, `zoom` (how far the picture punches in on a highlight, 1.2), `face_zoom` (the facecam's punch-in in `split`, 1.15), `wide_zoom` (punch-in on wide shots in `crop`, 1 to turn it off).
- - Key words: `highlight_font` and `highlight_scale` set highlights in a second font, larger, and `highlight_upper` puts them in capitals. `emphasis_font`, `emphasis_scale` and `emphasis_color` do the same for emphasis words, and default to the highlight ones.
- - Layout: `line_words` breaks a caption into lines of that many words, and a key word gets its own line. `stagger` steps each line right by that many pixels. `reveal: true` shows each word as it is spoken. `frame: "square"` puts a 1:1 picture between black bars, for the crop layout. `kinetic` (such as 1.35) draws any caption holding a key word that many times bigger, so plain lines stay subtitle-sized.
- - Hook: `hook_font`, `hook_highlight` (the starred word), `hook_box` (box colour).
- Fonts are macOS font names such as `HelveticaNeue-Medium`, `Didot-Bold` or `AvenirNext-Heavy`. On Windows and Linux the built-in styles use the closest fonts there, so they look slightly different. A style can also name a font file, such as `C:/Windows/Fonts/impact.ttf`.
- ## Teach it a style
- When the user shows a Short they like, look at several frames of it and write the closest matching keys to `~/.claude/skills/gta-clipper/styles/<name>.json`, with a short name the user picks. From then on it is one of the styles offered in step 3. Start from the existing style closest to it and change only what differs. Check for an outline or a shadow, capitals or lower case, words appearing one at a time, stepped lines, one or two accent fonts for key words, and a title box at the start. Copy the caption style only, never the video.
README.md · README · 74 lines
- # GTA Clipper
- A free Claude Code skill that turns a finished GTA 6 stream into vertical Shorts: the streamer's facecam on top, the gameplay below, captions in between, and zooms on the big moments. It picks the best moments, shows you a ranked list, and only cuts what you approve.
- Give it a Twitch VOD or clip, a Kick VOD, a YouTube link or a video file. For a link it downloads only the audio to find the moments, then only the video each approved clip needs, so a 14-hour stream doesn't fill your disk.
- Everything runs on your own computer except picking the moments, which uses your Claude plan.
- ## What you need
- - Claude Code, which needs a paid Claude plan.
- - A Whisper model, about 150 MB. The skill offers to download one if you don't have it.
- On a Mac:
- - FFmpeg, whisper.cpp and yt-dlp: `brew install ffmpeg whisper-cpp yt-dlp`
- - Swift, from the Xcode Command Line Tools: `xcode-select --install`
- On Windows:
- - Python 3 from python.org, then `pip install opencv-python-headless pillow`
- - FFmpeg and yt-dlp: `winget install Gyan.FFmpeg yt-dlp.yt-dlp`
- - whisper.cpp: download `whisper-bin-x64.zip` from the [whisper.cpp releases](https://github.com/ggml-org/whisper.cpp/releases), unzip it, and add the folder with `whisper-cli.exe` to your PATH.
- yt-dlp is only needed for Twitch, Kick and YouTube links.
- ## Install
- Paste this repo's link into Claude Code and say: "Install this skill and set it up for me."
- Or do it by hand:
- ```bash
- git clone https://github.com/Itsme23476/gta-clipper ~/.claude/skills/gta-clipper
- ```
- The skill must live in `~/.claude/skills/gta-clipper`. Restart Claude Code after installing. It's separate from [Claude Clipper](https://github.com/Itsme23476/claude-clipper), so you can have both.
- ## Use
- In Claude Code, type `/gta-clipper` and paste a finished stream's link, such as `https://www.twitch.tv/videos/<id>`. You get a ranked list of moments with a score, start time and opening line for each, and it asks which style you want. Tell it which clips to keep, and it renders them to `~/Desktop/Clips/<stream name>/clips/`, with titles and post captions in `post-copy.md`.
- To change anything, just ask: a bigger facecam, no zooms on one clip, a shorter cut.
- ## Layout
- - **Stream with a facecam on the left or right side:** the facecam on top, taking 40% of the height, and the gameplay below with 60%. Captions sit on the seam, and the gameplay zooms in on the big moments.
- - **Just Chatting or a talking head:** a crop that follows the face.
- - **No facecam:** the full frame, with bars above and below.
- ## Styles
- - **Hormozi** (default): bold white captions build word by word, punch words land in giant capitals and key words in red script.
- - **Elegant**: lower-case captions that build word by word in stepped lines, with key words in gold serif.
- - **Classic**: bold outlined capitals with yellow highlight words.
- - **Podcast**: small white subtitles that jump up in size on key words.
- Each style is a small file in `styles/`. Show it a Short whose captions you like and ask it to learn that style, and it saves a new one there that you can pick next time.
- ## Limits
- - It's tested on a Mac. On Windows it uses OpenCV and Pillow in place of the Mac's own frameworks. That path was tested on a Mac, not yet on a Windows PC. Captions use the closest Windows fonts, so they look slightly different.
- - Finished streams only, not live ones.
- - It picks moments from what's said, not what's on screen, so a moment that only looks good won't be found.
- - Subscriber-only VODs need your Twitch login. Download those yourself and give it the file.
- - Long streams take a while to download and transcribe.
- - It's set up for English. Whisper mishears names, so read the captions and tell it what to fix.
- - Loud game music can make Whisper repeat one line for minutes, and nothing said in that stretch can be clipped.
- Only clip videos you own or have permission to use, such as a streamer's clipping campaign. Read each campaign's rules before posting: some don't accept AI-edited clips.
- ## License
- MIT
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.