noface-clipper
Turns long gaming streams and videos (GTA 6 and any other game) into finished vertical 9:16 Shorts that show ONLY the gameplay, never the streamer's face. Takes a Twitch, YouTube or Kick VOD link or a video file, transcribes it on this computer, finds the best moments from chat spikes, voice peaks and viewer clips, hides the facecam, adds word-by-word captions, a hook, a cover and post text, runs a quality check, lets the user approve each Short, can add the user's own HeyGen avatar as a webcam under the game with an English commentary and captions, then can publish the approved ones with Blotato. Use it whenever the user wants to clip a stream, make Shorts, Reels or TikToks from a gameplay video or Twitch link, "make 5 Shorts from <link>", "clip this VOD", "gameplay only, no face", wants their avatar, face or voice commentary on the Shorts, or wants to review or post clips this skill made, even if they don't name the skill.
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/drfirass--gta6-noface-clipper)HTML
<a href="/listings/drfirass--gta6-noface-clipper"><img src="/api/listings/drfirass--gta6-noface-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.
Requirements: [Claude Code](https://claude.com/claude-code) (paid Claude plan), Python 3.10 to 3.12, git, and ffmpeg. On a Mac, [Homebrew](https://brew.sh).
**macOS / Linux**
git clone https://github.com/DrFIRASS/gta6-noface-clipper.git ~/.claude/skills/noface-clipper python3.12 ~/.claude/skills/noface-clipper/setup.py
**Windows (PowerShell)**
winget install Gyan.FFmpeg git clone https://github.com/DrFIRASS/gta6-noface-clipper.git "$env:USERPROFILE\.claude\skills\noface-clipper" py -3.12 "$env:USERPROFILE\.claude\skills\noface-clipper\setup.py"
`setup.py` creates a private Python environment for the skill, installs its libraries and checks ffmpeg. On a Mac whose ffmpeg has no subtitle filter, it installs Homebrew's `ffmpeg-full` next to your ffmpeg and uses it for the skill only. It takes a few minutes the first time. Restart Claude Code afterwards so it sees the new skill.
**Update:** `git -C ~/.claude/skills/noface-clipper pull`, then run `setup.py` again.
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 · 78 lines
- ---
- name: noface-clipper
- description: Turns long gaming streams and videos (GTA 6 and any other game) into finished vertical 9:16 Shorts that show ONLY the gameplay, never the streamer's face. Takes a Twitch, YouTube or Kick VOD link or a video file, transcribes it on this computer, finds the best moments from chat spikes, voice peaks and viewer clips, hides the facecam, adds word-by-word captions, a hook, a cover and post text, runs a quality check, lets the user approve each Short, can add the user's own HeyGen avatar as a webcam under the game with an English commentary and captions, then can publish the approved ones with Blotato. Use it whenever the user wants to clip a stream, make Shorts, Reels or TikToks from a gameplay video or Twitch link, "make 5 Shorts from <link>", "clip this VOD", "gameplay only, no face", wants their avatar, face or voice commentary on the Shorts, or wants to review or post clips this skill made, even if they don't name the skill.
- ---
- # noface-clipper
- A video editor for gameplay clips. It cuts a stream into short vertical clips with captions and a credit line, shows only the game (the facecam is hidden on every frame), and never touches the game itself.
- **Rights.** The user is responsible for having permission to clip and publish each streamer (direct permission or an official clipping program) and for the platforms' rules. Before the first run of a session, say this in one line, and check `SKILL/clip_rules.json` for the streamer (see step 1). Do not refuse a run over it: the decision is the user's.
- `SKILL` = this folder. `PY` = `SKILL/.venv/bin/python` (Windows: `SKILL\.venv\Scripts\python.exe`). Run from the user's folder; results go to `clips/`. Steps 1 and 3 take 10+ minutes: run them in the background and check the output. Missing `SKILL/.venv`: run `python3.12 SKILL/setup.py` (see Setup in `SKILL-REFERENCE.md`).
- Rare options (formats, caption styles, branding, memes, music, translation, live, voice, the app) are in `SKILL-REFERENCE.md`: read a section only when the user asks or a rule below sends you there. Never read whole transcripts: use the candidates list.
- ## 1. Get the source
- One link (the normal case):
- ```
- PY SKILL/watch.py <link> [START-END]
- ```
- Downloads audio only, writes `clips/src/<id>.txt` (sentences with timestamps and tags), `<id>.json` (word timings, needed to render), `<id>.meta.json` (title, channel, link). Streams over 4 h: pass a window of about 1 h (times stay full-video times). A list of sources the user chose can go in `SKILL/channels.txt` (empty by default), then `PY SKILL/watch.py "GTA ?(6|VI)" 5`.
- Tags say where to look, not what is good: `[LOUD +12dB]`, `[REPLAYED x5]`, `[CHAT x4 KEKW]` (chat spike), `[CLIPPED x2 5.6k views]` (viewers clipped it), `[MOMENT chase 0.8]`, `[HEAT 0-100]` (all signals in one number). Two tags on one line is usually the best moment.
- **Clip rules:** look the streamer up in `SKILL/clip_rules.json` (`platform:handle`, lowercase). `no` = tell the user they said no to clipping and go on only if the user says they have permission; `ask` = tell the user the note; `unknown` = remind them in one line to check the channel's rules.
- **Download fails:** tell the user to update yt-dlp with `PY -m pip install -U yt-dlp` and retry (do not run it silently). A bot or age check: wait 10 minutes, or ask the user before setting `CLIPPER_BROWSER=firefox` (it reads that browser's cookies).
- ## 2. Pick (your job)
- 1. `PY SKILL/nopick.py clips --candidates 30` (a few minutes on big transcripts: run it in the background). Read only `clips/candidates.txt`: the 30 best windows, one line each: `id start-end score tags | words (head ... tail)`.
- 2. Judge, rank, pick the best N (5 by default for a single link). To widen, shift or check a window, read only that region of `clips/src/<id>.txt` (grep the time range). Move a window's start or end to a sentence boundary; a story that needs more room gets it.
- 3. Before rendering run `PY SKILL/nopick.py --check clips/picks.json`: a window marked TOO SHORT ends up under 12 s. Widen it by a sentence or two, or pick another moment.
- 4. Fewer than N pass the rules: rerun with `--candidates 60`, then make fewer and say why. Skip ids in `clips/src/used.json` unless asked; never overlap ranges in `clips/src/moments.json`.
- No AI at all: `PY SKILL/nopick.py clips [N]` writes `picks.json` directly (weaker, no premise check).
- **A pick must have**
- 1. **Something happening in the game.** The Short shows only the gameplay, so the moment must be visible in the game: a chase, a crash, a fight, a stunt, a fail, a funny NPC, a big on-screen banner. The streamer's voice and chat are the signal and the soundtrack, not the picture. Talk about something off screen (backstory, opinions, sponsor reads, the chat, food) = 0.
- 2. **A premise and a payoff.** A peak (`[LOUD]`, `[CHAT]`, `[CLIPPED]`, `[REPLAYED]`, `[MOMENT chase/funny]`) and something a stranger can name said or shown before it. The payoff sits in the last 40 % and the Short ends right on it.
- 3. **Action first.** Second 1 shows the game in motion, not a menu, loading screen, map, pause screen or the streamer talking to the camera.
- 4. **No full-screen streamer.** Skip windows where the stream switches to a full-screen webcam or "just chatting" scene: those frames would be replaced by a frozen game frame.
- 5. **Only the game asked for** (GTA 6 unless another is named). **No spoilers or leaks:** no story endings, cutscene runs, leaked footage.
- 6. **Credit:** the streamer's or channel's real name (meta or words), else skip.
- 7. **Stands alone.** Never open on a pronoun with no antecedent, an answer without its question, or "and / but / so"; never cut mid-sentence. No two Shorts with the same point. Max 2 Shorts per source unless one link and N were asked.
- **Length follows the story:** one complete beat, start one sentence before the setup, never cut the payoff to hit a number. `batch.py` rejects a window over `MAX_SPAN`.
- **Text:** hook = 6 words max, caps, names the game moment (`COPS LOST HIM IN THE SWAMP`), says only what the clip shows. Title up to 70 chars; 3-5 hashtags; keywords = 1-3 words as spoken.
- Write `clips/picks.json`, best first (forward slashes):
- ```json
- [{"url": "<link from meta>", "transcript": "clips/src/<id>.json", "start": 2396.8, "end": 2421.6, "hook": "COPS LOST HIM IN THE SWAMP",
- "name": "swamp_chase", "score": 82, "title": "He lost five cop cars in the GTA 6 swamp", "hashtags": "#gta6 #gaming #shorts", "credit": "Streamer Name", "keywords": ["SWAMP", "COPS"]}]
- ```
- **Look first:** `PY SKILL/peek.py clips/picks.json clips/peek` saves 4 frames per moment as `clips/peek/<name>.jpg`. Open them. Drop a pick if a menu, map, ad, logo card or full-screen webcam fills the picture for more than a couple of seconds, or nothing moves in the game; fix a hook that mismatches the picture; replace and re-peek. Show the user the list (score, range, hook, why) unless told to just run.
- ## 3. Render
- ```
- PY SKILL/batch.py clips/picks.json clips/out
- ```
- Layout `nocam` is the default: the facecam is found and hidden, the game fills a large panel in the middle over a blurred copy of itself, the hook sits in the band above, captions on the lower part of the game. Downloads only each moment, adds zoom, captions and hook. Per Short: `NN_<name>.mp4`, `.txt` (title, hashtags, credit), `_cover.jpg`; `clips/out/overview.jpg` = 4 frames per Short. Several minutes per Short (`CLIPPER_WORKERS=1` on a weak PC). Another layout only if the user asks for the streamer to be shown (`CLIPPER_LAYOUT=stack`).
- ## 4. Check
- batch.py runs `qc.py` (PASS/WARN/FAIL, details `clips/out/qc.json`). FAIL: re-cut once with clip.py a sentence later or earlier as the reason says, re-run qc; still fails = drop it and say so. WARN `face visible`: open that Short's overview row and confirm it is a game character, not the streamer; if it is the streamer, drop the Short. Each `overview.jpg` row: first frame = the action (not a menu or empty street); last frame = still the moment; captions readable, hook readable, no black bars, no streamer face. Fix one alone: `PY SKILL/clip.py <url> <transcript> START END "HOOK" clips/out/<name>.mp4`, or drop it.
- ## 5. Hand over, approve, publish
- Give each Short with its `.txt` and the QC list. Offer the review app (`PY SKILL/app.py <the user's folder>` in the background, it opens the browser; see App in the reference) so the user approves (A) or rejects (X) each Short.
- Publishing happens only after the user approved the Shorts and named the account. With the Blotato MCP tools: list the accounts, confirm the target account and the privacy (`unlisted` or `private` for a test) with the user, upload each approved `.mp4`, then create the post with the title, description (title + credit line + hashtags) and the platform's required fields. Never post a Short the user did not approve. A Short with the user's avatar (step 6) is realistic AI content: YouTube `containsSyntheticMedia: true`, TikTok `isAiGenerated: true`, and tell the user to switch on Instagram's AI label in the app. Details: Post in the reference.
- ## 6. The user's avatar (optional, HeyGen)
- Only when the user asks for their avatar, face or voice on the Shorts. A second pass on Shorts already rendered (steps 1-4): no new VOD download, transcript or pick. The game moves to the top of the Short, their talking avatar fills the free band under it (a 760 px square with rounded corners), says an English commentary you write, and the captions (on the lower part of the game) become its words; the stream sound ducks under the voice. Details and settings: Avatar in the reference.
- 1. **Setup once.** The HeyGen MCP tools must be connected (custom connector `https://mcp.heygen.com/mcp/v1/`; if missing, give the user that step and stop). `~/.clipper/avatar.json` holds `{"avatar_id": "<look id>"}`. Missing: list their looks (`list_avatar_looks`, ownership private, digital_twin), show them (`show_avatars`) and save the one the user picks. Never pick, invent or swap the avatar yourself.
- 2. `PY SKILL/avatar.py plan clips/picks.json clips/out`: per Short its length, the word budget at this voice's real speed (target, min, max), which word lands on each zoom peak, and the streamer's lines with their times and word numbers. Writes `clips/avatar/scripts.json`. Only some Shorts (the user says "only one", "just the best"): copy those picks into `clips/picks_avatar.json` and use that file in steps 2-5 instead, since batch.py renders every pick in the file it gets.
- **Voice speed.** HeyGen voices speak 2.4 to 3.8 words per second, so a guessed speed can leave the Short with a frozen face for seconds or cut the speech. `get` measures the speed on every video and saves it in avatar.json, and plan and check use it. If plan says NOT MEASURED, generate the first Short alone (step 4), then run plan again and size the other scripts with the measured speed before generating them.
- 3. **Write the scripts** (English, `target_words` each, inside the min-max plan prints): first person, like a streamer reacting to their own run; short spoken sentences, present tense, contractions. Sentence 1 is the hook and lands in the first 2 s. Follow the clip's order using the word numbers plan prints, and put the reaction to the payoff on the peak's word. Say only what the game shows; no facts you can't see, no brackets, emoji, hashtags or stage directions. Put 1-3 of its words in `keywords`. Fill `script` and `keywords` in `scripts.json`, then run `PY SKILL/avatar.py check clips/picks.json clips/out` and fix every script it marks TOO SHORT or TOO LONG (it says how many words to add or cut and quotes the words on each peak). Then show the user every script with their credit balance (`get_current_user`) and wait for their OK: each video spends HeyGen credits.
- 4. **Generate**, one Short first: `create_video_from_avatar` with `avatarId` from avatar.json, `script` verbatim, `voiceId` from avatar.json if set (else the avatar's own voice, and then no `voiceSettings`, so the speed stays the one measured), `aspectRatio: "1:1"`, `resolution: "1080p"`, `fit: "cover"`, `title: "noface-clipper <name>"`. Poll `get_video` every 30 s, then `PY SKILL/avatar.py get <video_url> clips/avatar/<name>.mp4` (it saves the video and prints the voice speed it measured). Tell the user the credits the first one used. If the measured speed differs from the one plan used by more than 10 %, run plan and check again and adjust the remaining scripts before generating them. Never resubmit a video that is still processing.
- 5. `PY SKILL/avatar.py attach clips/picks.json`, then `PY SKILL/batch.py clips/picks.json clips/out_avatar` (the Shorts in `clips/out` stay). `AVATAR TOO LONG` / `AVATAR SHORT` in the log, or `still_s` over 1.5 s in the Short's `.render.json` (the face holds still at the end): shorten or lengthen that script with check and regenerate that one only, after the user's OK.
- Cover: batch.py makes `<name>_cover.jpg` with `cover.py`: the user's still photo in the avatar square and the game's name in a badge at the top, in the skill's font (`cover_label` in the pick or brand.json, default `GTA 6`; never the official logo). The photo is `~/.clipper/avatar_photo.jpg`: if missing, save the chosen look's `preview_image_url` there as a JPEG once.
- 6. Check `clips/out_avatar/overview.jpg` (game at the top, avatar square in the bottom band, English captions on the game, game in motion) and hand over as in step 5. QC ignores the face inside the avatar cam and still watches for the streamer.
README.md · README · 118 lines
- # GTA 6 No-Face Clipper
- **Type one sentence in Claude Code. Get finished 9:16 gameplay Shorts, with the streamer's face never shown.**
- `noface-clipper` is a free [Claude Code](https://claude.com/claude-code) skill. Give it a Twitch, YouTube or Kick VOD link (GTA 6 or any other game). It transcribes the stream on your computer, finds the moments where chat explodes and the voice peaks, hides the facecam, and renders vertical Shorts with word-by-word captions, a hook, a cover and ready-to-paste post text. You approve each Short, and it can post the approved ones to YouTube, TikTok and Instagram with [Blotato](https://blotato.com/?ref=firas).
- Optional: put **your own HeyGen avatar** on the Shorts, commenting the action in English, as if you were the one playing.
- > **Unofficial fan tool.** Not affiliated with or endorsed by Rockstar Games, Take-Two Interactive, Twitch, YouTube, Kick, TikTok, Instagram, HeyGen or Blotato. "GTA" and "Grand Theft Auto" are trademarks of Take-Two Interactive.
- > ## ⚠️ Your responsibility
- > noface-clipper is an editing tool. It does not give you any right to the videos you process. Streams, VODs and gameplay footage belong to their creators and to the game publisher. **You are solely responsible** for getting permission from the streamer (or joining an official clipping program) before you download, edit or publish any clip, for crediting them, and for respecting the terms of Twitch, YouTube, Kick, TikTok, Instagram and Rockstar Games / Take-Two. The authors of this skill accept no liability for how it is used.
- ## What you get
- - **Game only.** The facecam is detected and hidden on every frame. If the stream switches to a full-screen webcam, those frames are replaced by the last game frame. Big faces with no cam overlay are blurred.
- - **Built for Shorts.** The game fills a large panel over a blurred copy of itself, a hook (6 words max) sits at the top, word-by-word captions on the game, a short zoom on the peak.
- - **Ready to post.** Audio normalised to -14 LUFS, a cover image, a `.txt` with title, hashtags and the streamer's credit line.
- - **Quality check.** Every Short gets PASS / WARN / FAIL, including a "face visible" warning.
- - **Review app.** A local web page to watch every Short and approve (A) or reject (X) it.
- ## Optional: your own avatar on the Shorts
- If you have a [HeyGen](https://www.heygen.com) digital twin, the skill moves the game to the top of the Short and puts your talking avatar in the band under it (a large square with rounded corners). Claude writes an English commentary for each Short, sized to its length, and the captions switch to your words. The cover shows your photo and the game's name in a badge.
- 1. In Claude, add a custom connector with the URL `https://mcp.heygen.com/mcp/v1/` and connect your HeyGen account.
- 2. Ask: `add my HeyGen avatar to these Shorts`. Claude asks you to pick your avatar once, shows you the scripts and your credit balance, and generates nothing before you say yes.
- It is a second pass on Shorts already made: the stream is not downloaded or transcribed again. HeyGen videos use your HeyGen plan credits. Your avatar looks real, so label these Shorts as AI content on each platform (YouTube's "altered or synthetic content", TikTok's AI label, Instagram's AI info). The skill sets the flag on YouTube and TikTok when it posts.
- ## Install
- Requirements: [Claude Code](https://claude.com/claude-code) (paid Claude plan), Python 3.10 to 3.12, git, and ffmpeg. On a Mac, [Homebrew](https://brew.sh).
- **macOS / Linux**
- ```bash
- git clone https://github.com/DrFIRASS/gta6-noface-clipper.git ~/.claude/skills/noface-clipper
- python3.12 ~/.claude/skills/noface-clipper/setup.py
- ```
- **Windows (PowerShell)**
- ```powershell
- winget install Gyan.FFmpeg
- git clone https://github.com/DrFIRASS/gta6-noface-clipper.git "$env:USERPROFILE\.claude\skills\noface-clipper"
- py -3.12 "$env:USERPROFILE\.claude\skills\noface-clipper\setup.py"
- ```
- `setup.py` creates a private Python environment for the skill, installs its libraries and checks ffmpeg. On a Mac whose ffmpeg has no subtitle filter, it installs Homebrew's `ffmpeg-full` next to your ffmpeg and uses it for the skill only. It takes a few minutes the first time. Restart Claude Code afterwards so it sees the new skill.
- **Update:** `git -C ~/.claude/skills/noface-clipper pull`, then run `setup.py` again.
- ## Use
- Open Claude Code in an empty folder and type:
- ```
- Make 5 Shorts from https://www.twitch.tv/videos/1234567890
- ```
- Claude shows you the moments it picked (time range, hook, why) before rendering. The Shorts land in `clips/out/`. Then:
- ```
- open the clipper app
- ```
- to review them in your browser (A = approve, X = reject), and:
- ```
- add my HeyGen avatar to these Shorts
- ```
- for the optional avatar pass, and:
- ```
- post my approved Shorts with Blotato on my YouTube channel, unlisted
- ```
- to publish.
- ## Posting with Blotato
- The skill posts through [Blotato](https://blotato.com/?ref=firas), which connects YouTube, TikTok, Instagram and more in one place.
- 1. Create your account on [Blotato](https://blotato.com/?ref=firas) and connect your social accounts there.
- 2. Connect Blotato's MCP server in Claude (Settings, then Connectors).
- 3. Ask Claude to post. It lists your accounts, asks you to confirm the account and the privacy, and only posts the Shorts you approved.
- You can also upload the Shorts by hand for free: YouTube Studio, TikTok web upload, Meta Business Suite.
- ## How it finds the moments
- | Signal | What it measures |
- |---|---|
- | Chat spikes | Twitch chat replay in 10 s buckets, 2.5x the normal rate |
- | Voice peaks | 9 dB over the streamer's own normal level |
- | Viewer clips and replays | Moments viewers clipped or replayed |
- | Transcript | Whisper, word by word, on your computer |
- A local script ranks the 30 best windows. Claude reads only those 30 lines and picks the ones with something visible happening in the game and a clear setup and payoff that a stranger understands.
- ## Costs
- | Part | Cost |
- |---|---|
- | Download, transcription, detection, editing | Free, runs on your computer |
- | Claude Code | Paid Claude plan |
- | HeyGen (optional, for your avatar) | Your HeyGen plan credits |
- | [Blotato](https://blotato.com/?ref=firas) (optional, for auto-posting) | Paid plan; you can also upload by hand for free |
- ## Safety
- Compared to the original gta6-clipper: automatic updates are off, the review app no longer runs AI agents with permissions bypassed, autopilot scheduling is disabled, live AI modes are off, yt-dlp is never upgraded silently, links are passed safely to the downloader, and the Blotato key file is private (mode 600). Nothing is posted or generated with paid credits without your confirmation. See [CHANGELOG.md](CHANGELOG.md).
- ## Credits and license
- Made by Dr. Firas. Based on [gta6-clipper](https://gitlab.com/nocodehack.io/gta6-clipper) by nocodehack, MIT License. See [LICENSE](LICENSE) and [NOTICE.md](NOTICE.md).
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.