Cinematic
Reel
Drop in one raw vertical talking-head clip. Claude cuts the dead air, hides the bad takes and edits it into a premium investigative-documentary reel: letterbox, case-file HUD, evidence-board b-roll, a giant word behind your head and handwritten notes.
A full edit,
not a filter
The skill is a complete editing system. It has a locked visual style, a working template project and a step-by-step workflow, so every video comes out the same quality. You don't have to explain the style each time.
Upgrades the voice
It measures the raw audio for mud, hiss, clipping and dull tone, then applies a voice chain: light denoise, EQ (less mud, more clarity and air), de-esser, compression and −14 LUFS. For unusual problems it builds a custom chain and lets you A/B it. Before and after numbers are reported.
Removes retakes
Said a line three times? It rebuilds the script from the transcript and keeps the last clean take of each line. False starts, restarts and fillers like "umm" are cut.
Hides bad takes
It reviews every frame for closed eyes, looking away, odd faces and jump cuts. Each bad moment is covered by a full-screen evidence board while your voice keeps playing.
Removes silence
Dead air and long pauses go. Each cut starts just before the first word and ends where the silence begins, with short audio ramps so no cut clicks. The same word timings drive the subtitles and beats.
Documentary look
A natural skin grade, 2.39 letterbox moments, film grain, a typed FILE 01 HUD with running timecode and clean subtitles.
Motion graphics
Chapter titles with a red hand circle, an evidence board the camera flies across, and a giant word behind your head (AI person matte).
12 creative effects
Cover story, frame break, chart grow, vertigo, orbit, headline wall, paper cut, phone mockup and more, placed on your best face moments, 3–6 per reel.
Sound design
Whooshes on the letterbox, a camera flash on names, typing clicks, a stamp and a riser into the end, mixed under the voice with a master limiter so the voice never gets pushed down.
You approve first
Before anything renders, Claude shows a storyboard of every beat and waits for your OK. Render, loudness check and contact sheet come after.
Six beats, one story
Every reel is built from these six beats. Longer videos repeat them; see Beat modules.
Rack focus, slow push-in, the typed FILE 01 case label and running timecode.
A Kalam pen remark, written on, with a drawn arrow pointing at you.
Letterbox in, Anton words rise on the spoken words, a red circle and a Tillana reaction.
Full-screen b-roll that hides bad takes. The camera pans across clippings, polaroids and a stamp.
Red string to an index card, the key number circled and a hand note.
A giant word behind you via the person matte, a red rule and a riser.
Set up in
two minutes
Requirements
- Claude Code (CLI, desktop app or VS Code extension).
- Node.js 18+. HyperFrames runs through
npx hyperframes, so there is nothing else to install. - ffmpeg (
brew install ffmpeg). - macOS for source prep: it uses Apple's
avconvertto convert iPhone HDR (HLG) footage to standard video. On other systems, give it a clip that is already SDR. - Optional: an ElevenLabs account for fast transcription of clips over about 40 s. Local whisper works too, but it is slow on older Macs.
1 · Download and unzip
Download cinematic-reel-skill.zip. It contains one folder, cinematic-reel/. Or clone the GitHub repo and copy its cinematic-reel/ folder:
git clone https://github.com/aasim-shah/cinematic-reel-skill.git
cp -R cinematic-reel-skill/cinematic-reel ~/.claude/skills/2 · Put it where Claude can find it
For one project, put it in that project's skills folder:
mkdir -p .claude/skills
unzip ~/Downloads/cinematic-reel-skill.zip -d .claude/skillsOr put it in your personal skills folder to use it in every project:
mkdir -p ~/.claude/skills
unzip ~/Downloads/cinematic-reel-skill.zip -d ~/.claude/skills3 · Check it loaded
Start Claude Code in your video folder and type /cinematic-reel. If the skill appears in the list, you're ready.
/hyperframes, /hyperframes-core, /media-use) for the video engine. Install them first from the HyperFrames docs.One prompt
Put your raw clip in the project folder and ask Claude to edit it. Mention the clip with @ so Claude reads it.
@my-clip.mov edit this into a cinematic reelOther prompts that work:
/cinematic-reel @take-02.mov, the message is "AI won't take your job, people using AI will"
@raw.mov there are retakes and mistakes, cut them and hide bad faces with b-roll
make the storyboard for @interview.mp4 but don't render yetClaude asks only what it can't work out (for example the core message, if the speech is unclear). It then works through the steps below and stops at the storyboard for your approval. Reply "looks good, render it" or ask for changes.
What you get back
videos/<name>/renders/<name>-final.mp4: 1080×1920, 30 fps, about −14 LUFS, peaks below −1 dB.FACE-QC.md: every bad face moment found, and how it was hidden.STORYBOARD.mdandstoryboard/index.html: a visual card for every beat.- The full HyperFrames project, so you can tweak any beat and re-render.
Thirteen steps,
one gate
Claude follows the same pipeline every time. GPU-heavy jobs (matte, transcription, render) run one at a time.
videos/<name>/ and writes a brief.Locked tokens
The style lives in references/style.md. Claude never swaps colours, fonts or motion between videos, which is why every reel looks the same.
Palette: one accent only
Type
Grade
A natural skin grade. Faces look like the raw footage, not orange or crushed. The cinematic feel comes from the letterbox, HUD, grain and board, not from heavy colour. The exact grade JSON is in style.md.
Sound
| Moment | Sound | Rule |
|---|---|---|
| Typed label | click | One click per label, never per letter |
| Red circle lands | soft tick | One per title |
| Letterbox in / reopen | short whoosh | Soft, under the voice |
| Shutter to black, crash push | woosh | On the cut |
| Name or source appears, flash cut | camera flash | Faded out in 0.5 s |
| Stamp hits the board | mechanical click | One |
| Last line | sharp riser | Into the end |
Five building blocks
Every reel is made from these modules. Claude changes the words, the timing and the board contents, never the design.
| Module | On screen | Used for |
|---|---|---|
| 1 · Cold open | Rack focus, slow push-in, typed FILE 01 / TOPIC, timecode, subtitles, one pen note | The hook |
| 2 · Chapter title | Letterbox, two-line Anton title rising word by word, red circle on the keyword, red Tillana reaction | Each new idea |
| 3 · Shutter | Bars close to black, title rides up and blurs out | Transition into a board; hides a short bad take |
| 4 · Evidence board | Felt board, newspaper clipping, polaroids, chart, index card, stamp and red string, with the camera flying across | Names, sources, numbers; hides bad takes |
| 5 · Behind the head | Flash cut back, room dims, giant word behind the speaker, red rule, tag, hand note, riser | The key claim and the ending |
Longer videos (30–120 s)
- Order: cold open → (chapter title → board or face beat) × N → behind-head ending.
- About one beat every 3–6 seconds: roughly 8 beats for 30 s and 15 for 60 s.
- A new chapter title (and
FILE 02,FILE 03…) for each new idea, at most one every ~10 s. - One evidence board per bad take, and at least one every ~15 s, each labelled
EXHIBIT A,B,Cwith a different mix of items. - About one hand note per beat, never two on screen at once.
Twelve effects,
used sparingly
On top of the beat modules, Claude places 3–6 creative effects per reel, each on a good face moment that matches its line. At most one every 6–8 seconds: a few well-timed effects look premium, a parade looks like a template.
Demos are cut from the effects showcase. The speaker is shown as a dark silhouette here; in your reel it is you.
The package
cinematic-reel/
├── SKILL.md workflow, quick reference, common mistakes
├── references/
│ ├── style.md palette, type, grade JSON, motion, sound map
│ ├── modules.md the five beat modules and how to scale them
│ ├── effects.md the 12-effect catalog and integration rules
│ └── qc.md Audio QC, Face QC, recording tips, render notes
├── fx/ the 12 effects as sub-compositions, notes, matte tools
├── scripts/
│ ├── new-project.sh copies the template into videos/<name>/
│ ├── audio-qc.sh voice analysis report
│ └── prep-source.sh HDR → SDR, 1080×1920 at 30 fps, voice clean-up
└── template/ the approved sample as a working HyperFrames project
├── index.html main timeline (all modules)
├── compositions/board.html
├── storyboard-template.html
└── assets/ fonts, textures, sound effectsQuestions
Which languages does it support?
Any language the transcriber handles. It was built for Roman Urdu / Hindi mixed with English: subtitles are written in the speaker's language, and Urdu/Hindi is written in Roman script.
How long does a render take?
About 2–5 minutes for a 15–30 s reel on a 2019 Intel MacBook Pro, more for longer videos. The person matte is the slowest step (about 1.3 s of CPU per frame), so Claude only mattes the short clips that need it.
Can I change the colours or fonts?
Yes. Edit references/style.md and the template together. Keep exactly one accent colour, and keep the handwriting to two fonts: one pen hand and one display hand.
My clip has lots of mistakes and retakes. Is that OK?
Yes, that's what it's built for. Dead air and restarts are cut. Bad faces are hidden under evidence boards while the voice continues. If there are too many to hide cleanly, Claude tells you and suggests which takes to drop.
Does it use stock footage?
No. Every b-roll is drawn in HTML/SVG (clippings, cards, charts, silhouettes), so there are no licences to manage and nothing looks generic.
The render is very slow.
Make sure only one render is running, and render from the CLI, not the HyperFrames Studio. The Studio can fall back to software rendering.
The voice got quieter after render.
The renderer lowers the whole mix if any peak goes over −1 dBTP. The template's master bus (+1.5 dB gain, −3 dBTP limiter) prevents that. Check the render log for "lowered".