Clip Miner
Finds the best 20-60 s moments in a long video and turns each into a captioned vertical short.
- Best for: podcasts, livestreams, webinars, interviews, long YouTube videos
- Category: Podcast / repurposing
- Default format: 9:16 (change with
--format 9:16 | 1:1 | 4:5 | 16:9) - Needs camera footage: yes, a raw talking clip
- Text behind the speaker: no
Before you start
This style runs on the video-studio-core skill, which must be installed in the same skills folder.
CORE= thevideo-studio-corefolder next to this one (e.g.~/.claude/skills/video-studio-core).STYLE= this skill's folder (it holdstheme.jsonandtheme.css).- First time on a machine:
python CORE/scripts/check_env.py, and fix anything it reports. - If
COREis missing, tell the user to installvideo-studio-corefrom the same pack.
Make the edit
python CORE/scripts/run.py clip.mp4 --style STYLE --work edit-clip-miner --stop-after plan
# refine edit-*/plan.json (see below), then check stills:
python CORE/scripts/render.py edit-clip-miner/plan.json --frames 1,3,6,9
python CORE/scripts/render.py edit-clip-miner/plan.json
python CORE/scripts/finish.py edit-clip-miner/plan.json
Open contact_sheet.png and look at it before the full render. The final video is final.mp4 in the work folder.
Use --formats 1:1,16:9 on finish.py for extra sizes. Add music with "music": {"src": "track.mp3", "volume": 0.12} in the plan.
The look
- Each clip must start on a hook (a question, a bold claim, a number) and end on a payoff. Never start mid-thought.
- Vertical 9:16, face centred, big captions, a hook title for the first 2.5 s, progress bar.
- Name each output file after its hook.
Refine the plan
The draft from make_plan.py is automatic. Make it good:
- Read
words.jsonand rewrite the hook (first 2-3 s) into a short, specific promise. - Replace weak auto-picked key words with the words that carry the meaning (numbers, results, names, contrasts).
- Fix any mis-heard words in the
wordsarray (keep their times). - Workflow:
prep.py long.mp4 --work long->transcribe.py --work long->clip_candidates.py --work long --top 12. - Read the candidates and the transcript around them. Pick the best N, adjusting start/end to clean sentence boundaries. Show the user your picks with times and hooks before rendering if they are around.
- For each pick:
run.py long.mp4 --style CLIP_MINER_DIR --work clip01after trimming withprep.py long.mp4 --work clip01 --start S --end E(run the rest of the steps on clip01). - Any other style can be used for the clips: pass that style's folder instead.
- If the source is 16:9 with the speaker off-centre, set
speaker.focusX(0 = left, 1 = right). - Element types and fields:
CORE/references/plan-schema.md.
Check before delivering
- Contact sheet: text readable, nothing cut off, nothing covering the face unless intended, safe zones respected (9:16: keep text between 10% and 85% of the height).
- Watch for timing: overlays land with the spoken word, not after it.
ffprobethe final file: right size, duration and audio.- Tell the user what you cut (seconds removed), what the look adds, and where the file is.
Change the look
Colours and fonts live in theme.json (vars, fontsUrl) and theme.css. For a one-off change, edit plan.json (theme.vars, element style) instead of the skill files.
Example requests
- "Find the 8 best moments in this podcast and make shorts."
- "Clip-mine my 1-hour livestream into reels."
- "Give me 5 viral clips from this interview, score the hooks."