augchan42/inkstone / skills/video-to-shorts
video-to-shorts
VerifiedUse when turning a long video (interview, podcast, talk, lecture, webinar) into short vertical clips for YouTube Shorts / Reels / TikTok. Transcribes with word timings, scores the most engaging self-contained moments, then renders each short from a JSON edit list - hook first, a cut to whoever speaks or reacts, punch-ins, tightened pauses, word-timed captions with the active word highlighted, a name lower third, a held ending that fades to a logo end card with a ding, speed variants at -14 LUFS. Handles two speakers on one wide shot (mic-channel attribution). Cross-platform (macOS / Linux / Windows-WSL). Stops at a finished, upload-ready vertical mp4 - no account or upload CLI required.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 13.1 KB |
| references/retention-edit.md | markdown | 7.2 KB |
| references/two-speaker.md | markdown | 3.5 KB |
| scripts/build-short.pyexecutable | script | 13.8 KB |
references/two-speaker.md
View on GitHub# Two speakers, one camera
Most podcast footage is one static wide shot of two people. A 9:16 crop holds one face at a
time. The short is built from two (or more) crops of the same frame, and it cuts between
them the way a multi-camera edit would. This file covers how to decide **who is speaking
when** and how to frame each person.
## Framing
- Crop width = source height x 9/16, rounded to an even number (1216 on 4K, 608 on 1080p).
- Give each person a base crop, e.g. `C` and `A`. Also give each a punch-in about 1.2x
tighter, e.g. `C+` = 1012x1800 on 4K. Anchor the punch-in to the bottom of the base crop
(y = 2160 - 1800 = 360) so the face gets bigger without moving up out of frame. Keep
every number even.
- Test each crop at 4-6 times across the clip, in one frame strip (`hstack` of crops with
white `pad` separators). People lean. A crop that is right at one moment can cut off a
head ten seconds later. Move the offset until every sample works.
- Never guess the offset. Show the strip and let the owner pick.
## Who is speaking: use the mic channels, not the picture
A frame shows who is on screen, not who is talking. Sampled stills land in pauses and prove
nothing. Take attribution from the audio.
**Two lavalier mics recorded to L and R** (the usual two-mic podcast kit): each speaker is
2-7 dB louder on their own channel. Measure it:
```python
# per word (start, end from the word-timestamp JSON), on the UN-normalised stereo master
l = 10*log10(mean(L**2)+1); r = 10*log10(mean(R**2)+1)
who = "L-speaker" if l - r > 1.5 else "R-speaker" if r - l > 1.5 else "?"
```
Calibrate the sign and the size on a known monologue first. On our rig the difference was
+4 to +7 dB for the left speaker and -2.5 to -5 dB for the right speaker. A `?` means
overlap, a laugh or a very short word. For those, check 100 ms windows:
```
1246.5 L 67.2 R 60.9 +6.2 C <- a word from C that the transcript does not have
1247.5 L 65.6 R 69.1 -3.5 A
```
Keep the **un-normalised** master for this. A mono or summed delivery master has lost the
difference.
## Recover words the transcript missed or misplaced
Whisper on a mixed track merges an overlap into one speaker or drops a short reply.
Transcribe each channel on its own over the few seconds in question:
```bash
ffmpeg -ss T -t 5 -i master.mp4 -af "pan=mono|c0=c0" -ar 16000 snipL.wav # c1 for R
```
Then run faster-whisper with `word_timestamps=True` on each file. The result from 2026-09-15:
- The mixed transcript had "What part of France are you from? … Normandy." as one block.
- The per-channel pass showed the guest answering "Normandy" at 1246.6 s.
- The host repeated "Normandy" at 1247.4 s, as the lead-in to the joke.
Two speakers and two shots, where the mixed transcript had one. Put the corrected words
into the edit list as a patch, so the captions use them.
If faster-whisper on CUDA fails with `libcublas.so.12 not found`, run it with
`device="cpu", compute_type="int8"` for a clip this short.
## Cut plan
1. List the words with start, end and speaker label for the whole moment.
2. Put a cut in the gap before each change of speaker, including one-word interjections.
3. Add reaction shots and punch-ins (see `retention-edit.md`), and trim the long pauses.
4. Write the plan as an edit list: shots in output order, each a source in/out plus a crop
name. Shots that continue each other in the source join without a fade. Every other
join gets 30 ms audio fades.
The audio always follows the source continuously inside a run of shots. Changing the crop
never touches the sound.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |