talk-edit
/talk-edit
Montuje nagranie mówcy przed kamerą bez edytora osi czasu: wycina pauzy, wyrównuje głośność i dodaje napisy.
Ocena i bezpieczeństwo
- Pliki w paczce
- 7
- Skrypty
- 6
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 07.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 6 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/build.py, scripts/check_takes.py, scripts/cut.py, scripts/level_audio.py, scripts/make_srt.py, scripts/transcribe.py
Co robi i kiedy
Montuje nagranie mówcy przed kamerą bez edytora osi czasu: wycina pauzy, wyrównuje głośność i dodaje napisy.
Wyzwalacze z opisu
Oryginalny opis (EN)
Edit a talking-head or screen-recording video (a vlog, a tutorial, a confessional, a talk) from raw takes to a finished cut with no timeline editor - check the takes, transcribe every word with timestamps, plan the cut from the transcript, cut on real silences with air at every edge, tighten long pauses, cut out mic dropouts, crossfade the joins, centre and level the voice to -14 LUFS, blur private info, make subtitles, then refine it in review rounds on the review page. Use it when someone hands over recordings of themselves talking and wants an edited video, says "cut the ums", "tighten the
Jak działa
- A project folder
- Steps
- Privacy, before anyone sees a cut
- Mistakes on camera
- Graphics and extras (optional)
- Saying where the AI went
- Files
Podgląd SKILL.md
---
name: talk-edit
description: Edit a talking-head or screen-recording video (a vlog, a tutorial, a confessional, a talk) from raw takes to a finished cut with no timeline editor - check the takes, transcribe every word with timestamps, plan the cut from the transcript, cut on real silences with air at every edge, tighten long pauses, cut out mic dropouts, crossfade the joins, centre and level the voice to -14 LUFS, blur private info, make subtitles, then refine it in review rounds on the review page. Use it when someone hands over recordings of themselves talking and wants an edited video, says "cut the ums", "tighten the pauses", "edit my talk/vlog/tutorial", "make subtitles", or wants a video made "without Premiere".
---
# Talk edit
The person talks; Claude cuts. Every cut is still theirs to approve: they review drafts on the review page (the
**review-harness** skill in this repo) and their notes drive the next draft.
Needs: ffmpeg on the PATH, Python 3 with `numpy` and `faster-whisper` (`pip install numpy faster-whisper`).
## A project folder
```
takes.txt one line per recording: c1 C:/Videos/take1.mp4 (short names; the cut refers to them)
prompt.txt optional: names and words for whisper to spell right ("Claude, ffmpeg, AIVCR")
plan.json the cut: which pieces of which takes, in order (cut.py)
fixes.json optional: [regex, replacement] pairs for subtitles whisper still gets wrong
blur.json optional: boxes to blur, [take, t0, t1, x, y, w, h] in take seconds and pixels
```
## Steps
1. **Check the takes right after recording**, while a line can still be redone:
`python scripts/check_takes.py --takes takes.txt`. It lists mic dropouts (exit 1 with a REDO list), clicks, samples
at full scale (turn the gain down) and a lopsided left/right balance (fixed later, no redo).
2. **Transcribe:** `python scripts/transcribe.py` writes `words/<take>.json` (every word with its start and end) and
`transcript_<take>.txt` to read.
3. **Plan the cut with the person.** Read the transcript; propose what to keep and what to drop (false starts,
repeated lines, rambles, dead air), with the times. They decide. Write it into `plan.json`:
```json
{ "fps": 30, "min_pause": 1.3,
"pieces": [ {"take": "c1", "in": 0.5, "out": 94.8, "note": "the intro"},
{"take": "c1", "in": 103.8, "out": 573.6, "exact_out": true, "note": "keeps the gesture after the last word"} ] }
```
Uncut is a valid choice too (one piece, `"tighten": false`): some talks are about honesty and read worse cut.
4. **Cut:** `python scripts/cut.py` writes `timeline.json`. The rules it follows:
- edges only in real silence, found on the audio itself, never on whisper's word times (they run 0.1-0.2 s off, and a
cut placed from them clips the ends of words);
- 0.24 s of air after the last word and 0.16 s before the first, so cuts never feel abrupt;
- pauses of `min_pause` s or more are tightened, keeping 0.45 s after the speech and 0.40 s before the next (cutting
too close and too often is the commonest complaint);
- mic dropouts (runs of digital silence) are cut out;
- every piece is a whole number of frames, so anything laid over the picture stays in sync to the end.
5. **Build:** `python scripts/build.py` writes `edit.mp4`: the picture piece by piece (with `blur.json`'s boxes), the
sound with crossfaded joins over a faint loop of the room's own tone (no join drops to dead silence), then
`level_audio.py`: the voice folded to centred mono (some mics record one channel louder, which sounds lopsided),
a high-pass at 80 Hz, a limiter, and a single steady boost to -14 LUFS (YouTube's level). It stops rather than let
loudnorm ride the gain.
6. **Subtitles:** `python scripts/make_srt.py` writes `edit.srt` from the words that made the cut, in short lines.
7. **Review:** add the draft to the review page (`python ../review-harness/scripts/add_draft.py <project> edit.mp4`),
give the person the link, and wait for their notes. Read every note and its marked-up frame before changing anything.
Edit `plan.json`, re-run 4-6, add the next draft. The page shows what changed and where the work keeps returning.
## Privacy, before anyone sees a cut
Screen recordings leak: email addresses, address bars, private links, other people's messages. Look through every
screen-recorded stretch (pause on each screen change) and add a `blur.json` box for each. If a private link ever showed in
a take, treat it as public: revoke it.
## Mistakes on camera
Keep the take flowing. When the person states something wrong, the usual fix is to keep their line and add a short
on-screen correction (a caption, a small card), not a re-record. Fact-check claims with the person before recording.
## Graphics and extras (optional)
Cards, lower thirds, diagrams and screenshots go in the empty parts of the frame (a blank wall, a dark column beside a
screen share), never over the face or the hands, and never zoom in on someone who talks with their hands. Any tool works:
ask Claude to build them as code (Remotion, HTML rendered with Playwright) or as PNG/MOV files, keyed to times in
`timeline.json` (a take second maps to output time through its piece's `in` and `start`). Composite them over `edit.mp4`
with ffmpeg's `overlay`.
## Saying where the AI went
The video should say how it was made, on screen or in the description. The **slop-o-meter** skill in this repo makes a
card that scores it out of 100: a talk that is mostly the person on camera scores low.
## Files
- `scripts/check_takes.py`: takes report (dropouts, clicks, clipping, balance).
- `scripts/transcribe.py`: words with timestamps per take.
- `scripts/cut.py`: plan.json -> timeline.json (the edge rules above).
- `scripts/build.py`: timeline.json -> edit.mp4 (picture, crossfaded sound, levelling).
- `scripts/level_audio.py`: centred mono, -14 LUFS, linear or nothing.
- `scripts/make_srt.py`: subtitles.
Pliki w paczce (7)
- SKILL.md 6 KB
- scripts/build.py 6 KB
- scripts/check_takes.py 5 KB
- scripts/cut.py 7 KB
- scripts/level_audio.py 4 KB
- scripts/make_srt.py 2 KB
- scripts/transcribe.py 3 KB
Wersje
W bazie nie ma innych skilli o nazwie talk-edit.
Migawka bazy z 13.11.2026 · wersja paczki 51643c271e24 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100cut-silences
Usuwa ciszy i martwe pauzy z nagrania z mówiącą osobą, tworząc listę cięć i przesuniętą transkrypcję dla kolejnych etapów. - 100cut-silences
Usuwa ciszy i martwe pauzy z nagrania z mówiącą osobą, tworząc listę cięć i przesuniętą transkrypcję dla kolejnych etapów. - 100video-use
Montuje nagrania sterując rozmową: wycina pauzy i powtórki, przełącza kamerę i ekran oraz wypala napisy. - 90resolve-tighten-recording
Skraca jedno długie nagranie w DaVinci Resolve, usuwając martwe pauzy i ciszę, aby powstał pierwszy wariant montażu. - 100edit
Montuje filmy z mówiącą osobą od początku do końca: tnie ciszę i wpadki, dodaje napisy zgrane ze słowami, normalizuje dźwięk do -14 LUFS i eksportuje na Reels, TikTok i Shorts. - 100edit-footage
Zamienia własne surowe nagranie mówione w pionową rolkę lokalnie i bez kosztów: transkrypcja, wybór właściwego ujęcia, usuwanie pauz i napisy.