Skip to content

GuruDevvv/my-claude-skills / mine/reels-console

reels-console

Verified

Edit a vertical reel (Reels, Shorts, TikTok) out of one static take — entirely in the console, no CapCut. Cuts pauses by speech, slices a single fixed shot into different crop levels anchored to the face, adds push-ins and a motion-blurred punch zoom, drops in B-roll, burns three-level captions from speech recognition, ducks music under the voice, masters loudness to -14 LUFS, embeds a cover frame, and checks its own cuts and captions before delivery. Use when asked to «смонтируй ролик», «собери рилс», «сделай видео динамичным», «нарежь дубль», «поставь субтитры на видео», «убери паузы из видео», edit my talking head video, make this reel dynamic, cut the silences.

FileKindSize
README.mdmarkdown3.7 KB
SKILL.mdskill md14.9 KB
references/recipes.mdmarkdown24.9 KB
scripts/reel.pyscript51.9 KB
# reels-console

A vertical reel cut from one static take, entirely in the console. No CapCut, no timeline.

## When to use

You filmed yourself talking to the camera from one fixed position and now need a reel out of it:
pauses gone, the picture changing every couple of seconds, captions burned in, music sitting under
the voice. Also for the single steps on their own — stripping silences, adding word-timed subtitles.

## What it does

Finds the pauses by sound and cuts them, keeping 0.18 s so speech does not sound chopped. Runs
speech recognition for word-level timing and burns three levels of captions: small base lines, two
to four accent words in large type, one call-to-action word in colour. Locates the face, works out
**how far you can push in before the crown of the head gets cut**, and slices the take into crop
levels with the eyes on the upper third. Adds slow push-ins, a pull-out before the finish, and a
punch zoom with real motion blur — rendered at 180 fps and averaged six frames into one, not a blur
filter. Overlays B-roll, ducks music under the voice with a sidechain compressor, and ends by
laying twelve frames into one contact sheet so the result can be checked at a glance.

```bash
python scripts/reel.py take.mp4 --dry-run        # the edit plan, nothing rendered
python scripts/reel.py take.mp4 -o reel.mp4 --accent first main result --cta subscribe
```

## Not for

Choosing the take, deciding which word carries the sentence, or writing the script — that stays with
the human, or with the model running the skill. It also does not shoot B-roll that was never filmed:
one take gives one angle, and the second angle has to exist.

## Needs

`ffmpeg` and `ffprobe` on PATH, plus `pip install openai-whisper opencv-python insightface
onnxruntime` for captions and face anchoring. Without them it still runs — no captions, crop from
the centre of the frame. A Cyrillic-capable font for captions (Arial and Arial Black by default).

## Honest about maturity

Written and tested in one evening, on two sources: a real 52-second phone take (vertical 4K, front
camera) and a synthetic bench with B-roll and music. Multiple faces in frame, horizontal video and
languages other than Russian are untested — whisper handles them, but the caption styling was tuned
for Cyrillic. Long recordings will be slow: whisper on a CPU is the bottleneck, not ffmpeg.

The one thing worth knowing before you shoot: on that real take the face filled 49.5% of the frame
width, so there was nowhere to push in and the main technique never engaged. Stand 1.5–2 m from the
camera. The script measures this and warns before rendering.

`references/recipes.md` has every technique as a standalone ffmpeg command, with measurements and
the traps — why `crop` cannot change size over time, how the motion blur is actually made, why
`amix` without `normalize=0` quietly eats the voice.

---

🇷🇺 Монтаж вертикального ролика из одного статичного дубля целиком в консоли: вырезает паузы,
нарезает съёмку с одной точки на планы разной крупности с привязкой к лицу, ставит наезды и ударный
наезд со смазом движения, кладёт перебивки, рисует титры трёх уровней из распознавания речи,
подкладывает музыку с автоприглушением. Что показать и какое слово главное — решает человек.
Приёмы взяты из разбора видео ilioshots «Как снять динамичный REELS на 1 кв. метре».
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: 198kUpdated Claude CodeCodexVerifiedInstall
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: 196kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 196kUpdated Claude CodeCodexVerifiedInstall
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: 196kUpdated Claude CodeCodexVerifiedInstall