Sponsored Reel Editor

You are editing a paid deliverable for a brand. The bar is higher than an organic post: the client will review it frame by frame and send timestamped notes. Everything here comes from revisions that got signed off.

Input: $ARGUMENTS

Requirements

ffmpeg and ffprobe on the PATH
Python 3 with Pillow (captions), openai-whisper (word timestamps) and numpy (clip scoring): pip install pillow openai-whisper numpy
Fonts: Manrope Bold (body captions, covers Greek), Inter Display Black and Inter Display Black Italic (hook title), and a serif italic for the hook's tail line (Georgia Italic, or Noto Serif Italic, which also covers Greek). The scripts find them in <project>/fonts/, then the usual system and user font folders on macOS, Linux and Windows. Override any of them with FONT_SUB, FONT_HOOK, FONT_HOOK_ITALIC or FONT_SERIF set to a file path.
SFX: the shared pack at ${CLAUDE_SKILL_DIR}/../edit-reel/audios/, or your own audios/ folder next to the video (that one wins when present).

The rule that governs everything

Never re-time the voiceover. Recut the picture only and re-attach the original audio untouched. Every caption timing, SFX cue and overlay is anchored to VO word timestamps, so the moment the audio shifts, every one of them breaks.

Concretely: build a new video-only track, then -map 0:v -map 1:a from the original. Total duration must match to the frame.

Pipeline

All scripts take the project folder (the video's folder) as their first argument and read and write work/ and edited/ inside it.

1. Transcribe for timings

model = whisper.load_model("medium")
r = model.transcribe(path, language="en", word_timestamps=True)   # or "el"

Cache to work/vo.json. Every downstream timing comes from these word stamps.

Proofread the transcript. Whisper reliably mangles brand and tool names: clod for Claude, lightning for lighting, a lowercase brand heard as a common word, and a domain like brand.ai split into two tokens. Apply a fixes map that replaces text without touching timings, and glue tokens starting with . onto the previous word.

2. Recut the picture

python3 ${CLAUDE_SKILL_DIR}/lib/recut.py <project>. Lay segments end to end with ffmpeg concat. Each segment is (source, in, out, optional filter).

Swapping in a reshot clip: trim it to exactly the slot duration it replaces.
Need a closer view but no reshoot? Digital zoom is fine when the source is 4K and the output is 1080p, because a 1.45x crop of a 2160-wide frame is still above target resolution. Say so rather than pretending it was reshot.
Lifting a dark shot: eq=brightness=0.06:contrast=1.04:saturation=1.05. Measure before and after and report the actual change. A grade is not the same as adding a light, and if the client asked for ambience lighting, tell them it is an approximation.

Print each segment's realised duration and confirm the sum.

3. Captions

Two named styles ship with this skill.

Pill captions with a glowing hook title (lib/render_reel.py)

Body captions:

White text on a translucent black rounded pill (alpha ~150, radius ~22), never a hard outline
Manrope Bold ~58px at 1080 wide
High in frame, around y = 0.20, because the bottom third is usually a lit keyboard or busy UI
Grouped 5 words max, split on sentence ends, no orphan tail of 1-2 words

Hook title:

Pairs of lines per screen: white upright caps on top, gold italic caps underneath
Gold is (245, 192, 19), sampled from the reference look, not guessed
Inter Display Black and Inter Display Black Italic; the tail line switches to a serif italic
No outline. Each word carries a tight glow in its own colour: draw the text to a layer, blur ~9px, composite at ~85%
A word does not exist before it is spoken. It appears at its timestamp, resolving from a ~14px blur over ~0.13s. Do not pre-render upcoming words at low opacity, that reads as a mistake

Bilingual: native-language line with English underneath (lib/render_bilingual.py)

Native line (for example Greek) Manrope Bold ~60px, English translation ~40px in light grey (214, 214, 214) underneath, both with a thin black stroke
The hook is carried by the subtitle itself, set in a heavier face (Inter Display Black ~70px) for the first few seconds, not by a second block of text saying the same thing
On sections where a UI fills the frame, drop the English so the product stays visible
The same script also does static punch-ins, pop-in image overlays, styled chat bubbles and the measured SFX mix

4. Zooms

Use static punch-ins, not animated ramps: hard cut into a tighter framing, hold, hard cut out. Values 1.08 to 1.12.

Align every cut in and out to a caption boundary or a real scene cut. A reframe landing mid-sentence looks like an error.

Find real cuts with:

ffmpeg -i in.mp4 -vf "select='gt(scene,0.30)',metadata=print" -an -f null -

5. SFX

This is where guessing fails. Measure, never assume.

Sounds come from ${CLAUDE_SKILL_DIR}/../edit-reel/audios/ (the shared pack) or an audios/ folder next to the video. The shared pack has trimmed_whoosh, trimmed_cash, trimmed_fah, trimmed_click, trimmed_bubble_pop, trimmed_riser and generated sfx_pop, sfx_swish, sfx_thock, sfx_thud, sfx_subdrop, sfx_shimmer, sfx_ding, sfx_buzzer, sfx_fanfare, sfx_tick (all .mp3).

Find where the sound actually starts. Stock clips often have a second of silence first. One clip had its click at 1.02s, so the naive first-half-second trim was pure silence at -91 dB.
```
ffmpeg -i sfx.mp3 -af "silencedetect=noise=-45dB:d=0.08" -f null -
```
Measure the trimmed slice with volumedetect and compute the gain needed to hit a target relative to the voice. With a VO around -18 dB RMS, transients want about -22 dB, beds about -28 dB, and an opening riser about -26 dB.
Apply volume={gain}dB, then alimiter=limit=0.97 on the final mix.

Rules: never the same sound twice in one video, place them on real scene cuts, and cap at four or five. Verify by comparing RMS of the final against the source over each SFX window; a cue that adds less than ~0.5 dB is inaudible.

6. Blur patches

For a notification, a face or a logo you cannot show. Crop the region, Gaussian blur it, and composite through a feathered mask so there is no visible rectangle.

Prefer this over recropping. A crop tight enough to lose a notification can also clip the client's own headline mid-word, which looks worse than the notification did.

For photos with bystanders, a graduated blur across the whole upper band hides every distant face and any screen content in one move and reads as depth of field rather than censorship.

Choosing clips by measurement

When handed long footage, do not eyeball it. Score every candidate window on motion and colour, drop anything too dark, and take the winner:

motion = np.abs(np.diff(frames, axis=0)).mean(axis=(1,2,3))
sat    = (frames.max(axis=3) - frames.min(axis=3)).mean(axis=(1,2))
score  = 0.6 * motion/motion.max() + 0.4 * sat/sat.max()

On 93 seconds of game footage this picked the one stretch that opened on a neon map and cut into a colour change. Tell the client how it was chosen; it reads as craft, not luck.

Filming conventions these edits assume

Phone pointed at the laptop screen, not a clean screen capture. It reads as someone showing you a thing rather than an ad
Browser zoom at 150% minimum. The single most common client note is that on-screen text is unreadable
Hide the bookmarks bar, close unrelated tabs, dismiss notifications
Dark room, screen as the brightest thing in frame, one warm lamp on the wall behind
Film 2 to 3x the seconds you need

Verify before you deliver

Never claim something is fixed without checking it:

Frame grids at 1s intervals to review structure: fps=1,scale=180:-1,tile=10x3
Tight crops upscaled with lanczos to judge text rendering
Brightness before and after any grade, as numbers
Audio RMS of final vs source per SFX window
Duration matches the source to the frame

What ships with this skill

lib/recut.py              picture-only recut, audio re-attached untouched
lib/render_reel.py        pill captions + glowing hook title
lib/render_bilingual.py   native-language captions with English underneath, punch-ins, overlays, SFX mix
lib/fonts.py              font lookup (project fonts/, system folders, env overrides)

Run each as python3 ${CLAUDE_SKILL_DIR}/lib/<script>.py <project-dir>. The timings, hook words and segment lists at the top of each file are a worked example; replace them with the values for your video before running. They are working code, not pseudocode.

Options

--greek transcribe with language="el" (any Whisper language code works the same way)
--bilingual English translation under the native-language line, ~40px in light grey, with the native line at ~60px above it. On sections where a UI fills the frame, drop the English so the product stays visible
--vo <file> when the voiceover was recorded separately

What not to do

Do not re-time or re-record the VO to fit the picture
Do not use animated zoom ramps
Do not put a black outline on the hook title
Do not show words before they are spoken
Do not repeat an SFX file
Do not guess SFX volumes as bare multipliers
Do not publish real DM screenshots or anything that fabricates evidence; use styled text bubbles instead

sponsored-reel

Files5

SKILL.mdplugins/edit-reel/skills/sponsored-reel/SKILL.md