clipper
Skill12 starsChanged 4 days agoMITView on GitHub
Turn a long podcast, interview, panel, livestream or X broadcast into a pack of topical, post-ready short clips, like an AI clipping SaaS but running on your own machine (macOS, Windows or Linux) with no per-minute fees. Downloads the source (X, YouTube, any yt-dlp link, or a local file), transcribes it on-device with Whisper (MLX on Apple Silicon, faster-whisper elsewhere), picks the strongest moments, cuts clean clips on word boundaries, QA-checks every cut, packages everything into a folder (plus zip) on the Desktop, uploads the videos to Google Drive with public links, and builds a shareable Google Doc with captions, timestamps and a link for every clip. Use when the user says "clip this", "make clips from this interview", "turn this podcast into shorts", "cut eli's podcast into clips", "the clipping thing", "clip it and put it on drive", "make the clips doc", or pastes a video link or file and wants clips.
4 other projects use a version of this.
Adopted by
Most recently taken up by
- Shash-Builds/Shash.buildstheir own version, first seen Nov 16, 2026
- jsleemaster/gooblintheir own version, first seen Nov 1, 2026
- Founderwithabs/SOLIDSFORYOUtheir own version, first seen Oct 25, 2026
- aj47/agent-skillstheir own version, first seen Jan 2, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- jsleemaster/gooblin★ 3same name
- aj47/agent-skills★ 1same name
- Founderwithabs/SOLIDSFORYOU★ 1same name
- Shash-Builds/Shash.builds★ 1same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 4 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 62
Read all of transcript.txt.
- Distinctive: line 94
Start on the first word of a complete thought (the question can be dropped if the answer stands alone).
- Distinctive: line 95
Prefer moments with a quotable line, a strong claim, a number, or a clean analogy.
- Distinctive: line 102
all lowercase, casual, the user's voice.
- Distinctive: line 104
No em dashes or en dashes anywhere.
- Very specific: line 131
Update start/end in plan.json, cut.py --only <name>, then check.py --only <name>.
- Very specific: line 167
Before handing over the link, verify what a reader gets: export the doc as text (https://docs.google.com/document/d/<id>/export?format=txt) …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
clipper
Long video in, post-ready clip pack out. Built for the Eli Ben-Sasson clip packs (The Block, Oct 2026, and "What If Bitcoin Can't Survive Quantum?", Nov 2026) and works on any long-form talk: video podcasts, interviews, panels and X broadcasts.
source (url or file)
-> fetch.py download to <work>/source.mp4
-> transcribe.py full.wav, segments.json, transcript.txt
-> YOU read the transcript, write plan.json (clips + captions)
-> cut.py clips/*.mp4 (and full_episode.mp4)
-> check.py head/tail transcripts, first/last frames, contact sheet
-> words.py word-level timing to fix any bad boundary, then re-cut
-> package.py ~/Desktop/<slug>/ + ~/Desktop/<slug>.zip
-> drive_upload.py Drive folder, every video shared "anyone with link"
-> make_doc.py Google Doc with captions + links, shared, link printedS below means this skill's scripts/ directory. Every script takes the work dir as its first argument. Work dirs live in ~/liveclip/<slug>/.
0. setup (once per machine)
PY=$(python3 S/setup.py) # prints a python with the whisper backend; reuses ~/liveclip/.venv if presentNeeds ffmpeg, yt-dlp, and gcloud for Drive. The whisper backend depends on the machine:
- Apple Silicon: MLX Whisper,
mlx-community/whisper-large-v3-turbo. - Windows, Linux, Intel Mac: faster-whisper,
large-v3-turbo. It uses an NVIDIA GPU when one is available (on Windows, setup.py installs the CUDA 12 cuBLAS and cuDNN 9 wheels), otherwise the CPU (int8, slower but fine for a single episode). - Overrides:
CLIPPER_BACKEND=mlx|faster-whisper,CLIPPER_DEVICE=cpu|cuda,CLIPPER_MODEL=<faster-whisper model>.
The model downloads on first run (about 1.6 GB).
On Windows, run the same steps from PowerShell: $PY = python S/setup.py, then & $PY -I S/transcribe.py <work> and so on. Use python wherever this file says python3.
1. fetch
python3 -I S/fetch.py "<url or path>" ~/liveclip/<slug>- Local file: symlinked as
source.<ext>, nothing copied. Where symlinks are not allowed (Windows without Developer Mode) it is hard-linked, or used in place fromsource_ref.txtif it sits on another drive. - URL: yt-dlp, best mp4. If X or YouTube wants a login, rerun with
CLIPPER_COOKIES=chrome. - X live broadcasts: the replay URL (
x.com/i/broadcasts/...) works once the stream ends. For a stream that is still live (macOS only for now), useS/live/liveclip <url>(rolling 20 min buffer into~/liveclip/buf) andS/live/clip <seconds>to grab the last N seconds. - Prints duration, resolution and size. Sanity check it is the right video.
2. transcribe
$PY -I S/transcribe.py ~/liveclip/<slug>Writes segments.json and transcript.txt with lines like [12:34 754.2-760.9] text. A 50 min episode takes a few minutes.
3. plan (this is the judgment step, do it yourself)
Read all of transcript.txt. Then write <work>/plan.json:
{
"slug": "eli-theblock-oct26",
"title": "eli ben-sasson on the starting block (the block), oct 26",
"source_url": "https://x.com/TheBlockCo/status/...",
"speaker": "eli ben-sasson",
"handles": "w/ @elibensasson @theblockco",
"pad": [0.3, 0.8],
"episode": {
"trim_start": 29.0,
"trim_end": null,
"caption": "eli ben-sasson joins ... \n\n\"verbatim quote\"\n\nfull interview below, timestamps in the replies\n\nw/ @elibensasson",
"timestamps": [["0:00", "intro"], ["5:26", "why starknet is researching its own l1"]]
},
"clips": [
{
"name": "01_l1_vote_93pct",
"title": "the 93% l1 vote",
"start": 379.04,
"end": 433.85,
"hook": "one line on what the clip is about",
"caption": "x-ready post. context line\n\n\"verbatim quote from the clip\"\n\nw/ @elibensasson @theblockco"
}
]
}How to pick clips:
- One idea per clip, 30 s to 2 min. Usually 6 to 13 clips per hour of footage.
- Start on the first word of a complete thought (the question can be dropped if the answer stands alone). End on the full stop of the answer, never mid-sentence or on "and so...".
- Prefer moments with a quotable line, a strong claim, a number, or a clean analogy. Include the episode's own cold open as clip 01 if it has one.
nameisNN_snake_case_topic. Order by how strong the clip is, or by episode order if the user prefers.start/endare seconds in the source.padadds pre-roll and tail at cut time (default 0.3 s and 0.8 s), so put the times on the words themselves.episodeis optional. Set it to cut a trimmed full episode (drops pre-show dead air viatrim_start). Timestamps are relative to the trimmed episode. Set"episode": nullto skip it.
Caption rules (default house style; if the user gives their own voice or style guide, follow theirs instead, but the verbatim-quote and no-dash rules always hold):
- all lowercase, casual, the user's voice. No emojis, no hashtags, no corporate framing.
- Quotes are verbatim from the transcript (lowercased). Never invent or "clean up" a quote beyond dropping ums.
- No em dashes or en dashes anywhere. Use commas, colons, periods. The scripts assert this and will refuse to write.
- Tag the speaker and the show at the end (
w/ @handle @show). Look up the show's real X handle if unsure. - Flag profanity in the hook (for example "contains the f-word, check before posting from a brand account").
4. cut
$PY -I S/cut.py ~/liveclip/<slug> # all clips + full episode
$PY -I S/cut.py ~/liveclip/<slug> --only 03_x # re-cut one clip after fixing plan.jsonH.264 CRF 18, AAC 160k, +faststart, short audio fades so cuts never click. Runs 3 encodes in parallel.
5. check (always, before showing anyone)
$PY -I S/check.py ~/liveclip/<slug>Prints every clip's duration plus what is said in its first and last 5 s, and writes first/last frames and check/contact.jpg. Read the output and open the contact sheet. A clip passes when the head starts on a clean first word, and the tail ends on a finished sentence with no next speaker sneaking in.
To fix a boundary, get word timings around it:
$PY -I S/words.py ~/liveclip/<slug> 2604 10 # words between 2604 s and 2614 s, absolute timesUpdate start/end in plan.json, cut.py --only <name>, then check.py --only <name>. Repeat until every clip passes.
6. package to the desktop
$PY -I S/package.py ~/liveclip/<slug> # -> ~/Desktop/<slug>/ and ~/Desktop/<slug>.zip
$PY -I S/package.py ~/liveclip/<slug> --dest ~/DownloadsThe folder holds every clip, the full episode, transcript.txt and clips.txt (source times, lengths, hooks, captions). Use the folder name the user asks for (--name "some folder").
7. upload to drive
Drive needs a gcloud login with Drive scope. The script checks first and stops with the fix if the token is missing or lacks Drive. Tell the user to run:
! gcloud auth login --enable-gdrive-accessUploading makes the files public ("anyone with the link"). Confirm with the user before the first upload of a new pack.
python3 -I S/drive_upload.py ~/liveclip/<slug>Creates a Drive folder named after title, uploads the full episode and every clip (resumable, 32 MB chunks), shares each publicly and records links in <work>/links.json. Safe to rerun: files already in links.json are skipped.
8. the doc
python3 -I S/make_doc.py ~/liveclip/<slug> --no-upload # writes doc.html only, read it first
python3 -I S/make_doc.py ~/liveclip/<slug> # creates the Google Doc in the folder, shares it, prints the linkThe doc has: title, Drive folder link, original source link, the full episode (link, length, caption, timestamps), then one section per clip (title, length, where it sits in the episode, caption, link).
Before handing over the link, verify what a reader gets: export the doc as text (https://docs.google.com/document/d/<id>/export?format=txt) without auth and check that the links work anonymously and that no markup leaked through.
9. hand over
Reply with: the Desktop folder path, the zip path, the Doc link, the Drive folder link, and the clip list (name, length, one-line hook). Note anything the user should check (profanity, third-party footage rights).
Origin
Built from real clip packs of Eli Ben-Sasson (StarkWare) interviews: The Block (8 clips plus full episode) and "What If Bitcoin Can't Survive Quantum?" (13 clips). Work dirs for past runs live under ~/liveclip/<slug>/ on the machine that made them.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

