video-director
Skill1 starsChanged 41 days agoMITView on GitHub
Director-level video production for any product, built in Remotion. Plans and makes launch and promo films, explainers, product demos, motion graphics over talking-head footage, captions and b-roll, long-to-short social cuts (Reels, Shorts, TikTok, LinkedIn), logo stings and music-driven montages, owning the whole pipeline from brief and storyboard through motion design, footage editing, QA, render and delivery, and orchestrating the scriptwriting, voice-direction and image-generation skills, with sound from the cinetic skill's score workflow. Reads the official Remotion skills on demand for the API. Use it whenever the user wants to make, edit, recut, caption or improve a video or animation, mentions Remotion, motion graphics, a promo, launch film, explainer, demo video or captions, or says a video looks generic, amateur or "AI-made", even if they never say "director".
10 other projects use a version of this.
Adopted by
Most recently taken up by
- kjranyone/artcraft-agent-skillstheir own version, first seen Nov 15, 2026
- mojomast/open-higgsfield-skillstheir own version, first seen Nov 15, 2026
- zuckenbert/video-from-codetheir own version, first seen Nov 8, 2026
- SallyRS/video-directortheir own version, first seen Nov 8, 2026
- uniqaiapp/video-directortheir own version, first seen Nov 1, 2026
- win4r/video-directortheir own version, first seen Nov 1, 2026
Other versions10 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 10
- win4r/video-director★ 400same name
- SallyRS/video-director★ 28same name
- mojomast/open-higgsfield-skills★ 8same name
- DotanVG/video-director-skill★ 2same name
- jayco2610/claude-skills★ 1same name
- kjranyone/artcraft-agent-skills★ 1same name
- sidgaikwad/skills★ 1same name
- uniqaiapp/video-director★ 0same name
- vabidev/AIVEC-Studio★ 0same name
- zuckenbert/video-from-code★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository4 other files it gives its agents.
- image-generationSkill
- scriptwritingSkill
- sound-designSkill
- voice-directionSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
13 rules. 0 are common, 10 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 8 edited versions we read keep these.
10
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 20
When this skill and an official Remotion skill disagree on an API fact, the official skill wins.
- Very specific: line 25
Read this SKILL.md in full first.
- Very specific: line 52
Before planning, read in the product repo: PRODUCT.md / BRAND.md / DESIGN.md (or frame.md / design.md), any audio or voice docs (for …
- Distinctive: line 99
Keep scope exact: never add scenes, narration, music or captions nobody asked for.
- Distinctive: line 113
Music sourcing: Search online libraries in this order: Pixabay Music, Freesound (filter licence CC0), Mixkit, Uppbeat.
- Distinctive: line 118
Prefer commercial use without attribution (Pixabay Content License, CC0, Mixkit free licence).
- Distinctive: line 120
Never NC (non-commercial) or ND (no derivatives).
- Distinctive: line 121
Never in-app platform music (Instagram, TikTok libraries): business accounts can't use most of it and the publishing API can't add it.
- Distinctive: line 127
One message, written as a claim: Every beat's job traces back to it (the feeling counts).
- Distinctive: line 146
Fill time with story, not wobble: No idle breathing or floating on content.
- Distinctive: line 151
Sound is half the film: Every visible state change that should sound gets one, placed by the cinetic score on the frame the picture …
- Distinctive: line 159
No AI mannerisms: No explanatory text the picture already says, no labels narrating visuals, no summary cards, no contrast reveals ("it's …
- Very specific: line 200
Start a new film by copying assets/project-template/ into the product repo as video/ (its own pnpm package, excluded from the host app's …
“Common” is measured: kept by at least 6 in 10 of the 8 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Video Director
You are the director, art director and editor. Remotion is the camera. This skill owns the film: what it says, how it looks and moves, how footage is cut, how it is reviewed and delivered. Specialist skills own their crafts and are called at the steps below; the official Remotion skills own the API.
| Skill | Owns | Called at |
scriptwriting | structure, hooks, writing for the ear, AI-sounding tells | Script |
voice-direction | casting, direction, accents, recording, take checks, VO build | Voice |
image-generation | illustrations, characters, plates, keyframes, cutouts | Style frames, Build |
cinetic (sound only) | the score: earshot SFX kit, cues, music bed, mix and master (its references/sound.md) | Sound |
remotion-best-practices (official) | Remotion API, Studio, rendering | when code needs an API |
When this skill and an official Remotion skill disagree on an API fact, the official skill wins; on taste or process, this skill wins (see "Where we deviate"). Product docs (brand, voice, copy rules) override every skill's defaults.
Step 0: read what the step needs, when it needs it
Read this SKILL.md in full first. Read a reference in full when you reach the step it serves; don't skim a file you
rely on, and don't load files the task never touches.
This skill's references, by step:
- Brief and script:
references/direction.md(brief, pitch round, story, pacing, storyboard, honesty). - Style frames:
references/design.md(layout, type, colour, brand, real product screens, formats). - Build:
references/motion.md(easing, springs, kinetic type, seams, camera) andreferences/pipeline.md(project structure, timeline, builders, assets, Remotion file map);references/footage.mdwhen there is footage or captions;assets/project-template/README.mdfor a new project. - Review:
references/qa.md(review sheets, the single review, the restate test, the anti-AI pass) andassets/CRITIC_BRIEF.template.md. - Deliver:
references/delivery.md(encode settings, poster, platform exports, hand-off). - Scripts:
scripts/README.mdwhen you run one.
The social track reads only the project's
CREATIVE.mdandREADME.md, the safe-zone and caption sections offootage.md, andqa.md§3 and §7. The project's code already carries the rest.- Brief and script:
- Official Remotion skills, on demand. Invoke
remotion-best-practicesand follow its "if X, load Y" routing: read the files for the APIs the code you are about to write uses (the map inpipeline.md§8 finds them). Check the project's Remotion version against theversion:line; useremotion-upgradeor adapt if an API is missing. Never recall Remotion APIs from memory. - Specialist skills. At their step, invoke the skill named in the table above (or read its
SKILL.mdin the sibling folder). Installed as a plugin, the names carry a prefix (media-production:scriptwriting). For sound, read the installedcineticskill'sreferences/sound.mdand use only its sound workflow. Don't re-derive their rules here.
Step 1: read the product, then pick the mode
Before planning, read in the product repo: PRODUCT.md / BRAND.md / DESIGN.md (or frame.md / design.md), any
audio or voice docs (for example AUDIO.md), illustration docs, the real UI (run the app or capture screenshots),
pricing and claims in code, and an existing video/ package. No design doc: capture the live product and extract tokens
before designing anything.
| Mode | Typical ask | Clock | References to read |
| Launch / promo film | "make a launch video", "sell it from the pain" | VO if narrated, else music | direction, design, motion, qa, delivery |
| Explainer | "explain how X works", article or doc to video | VO | direction, design, motion |
| Product demo | "show the flow", screen recording polish | VO or cursor beats | footage (screen demos), design (real product), motion |
| Talking-head overlay / recut | "add graphics to this interview", b-roll, cards | the speech | footage, design |
| Captions only | "caption this clip", word pop, karaoke | the speech | footage (captions) |
| Long to short | "cut this podcast into Shorts" | the speech | footage (long to short), scriptwriting hooks |
| Music montage / sting | "beat-synced reel", "logo animation" | music | motion, direction (music pacing), cinetic sound |
| Remake a reference | "make ours like this video" | the reference | direction (reference-first), qa (reference compare) |
| Social cut | "3 Reels this week", a hook test, a card demo, a recut of the film | VO or music already in the project | social track below |
Two tracks
Film track: a new film, a new look, or a project without a locked style.
Social track: short cuts (6 to 45 s) built in an existing project whose brand, motion library, voices, music and
illustrations are already locked, usually a batch of 3 to 5. Reuse the locked assets; write each cut with the
scriptwriting short-form path; record new VO lines with the locked voice (no casting); build each cut as a new
composition from the project's components; score it; render; then one review for the whole batch.
Step 2: make the film
The steps run straight through. There are no gates between them: stop for approval only when the user asks for it in an interactive session. Autonomous runs make visible decisions with a one-line reason each in the decision log.
- Brief. One-sentence message written as a claim; audience; destination (muted feed or sound-on) and formats;
length; reference style named ("Linear launch", not "premium modern"); facts file. Vague asks get the five-concept
pitch round first (
direction.md). - Script, storyboard, beat map.
scriptwritingwrites it; durations add up; every animated event that should sound is marked. - Voice.
voice-direction: cast, record, build loudness-matched lines with word timings; captions follow what the audio says. - Style frames. Stills with real copy, fonts and colours; generated assets through
image-generation. The build dresses this layout; it never redraws it. - Build. Scenes from the shared core (
pipeline.md); types pass. - Sound. The cinetic score workflow plus a music bed found online (see "Sound" below).
- Render. Draft at low resolution; the render and deliver scripts measure spec, loudness and sync as part of the render, not as a review step.
- Review, once. Three critics, fixes in, done (
qa.md§3). - Deliver. Master, platform exports, poster, hand-off.
Keep scope exact: never add scenes, narration, music or captions nobody asked for; offer them instead. SFX on visible events are part of the craft, not an addition: they are on unless the brief asks for silence.
Sound
For video, sound always comes from the cinetic skill's workflow (references/sound.md; scripts scripts/audio/kit.py,
score.py, listen.py): an earshot SFX kit placed on cues exported from the picture code, music as the score's bed
track, mixed and mastered by score.py. The project's timeline exports the cues; earshot lives in its own virtualenv
(EARSHOT_PYTHON). Don't call sound-design for video.
With a voiceover. score.py has no voice track. Render its soundtrack (music and effects) as a stem, then lay the
VO over it with a sidechain duck (ffmpeg sidechaincompress, music about 8 to 10 dB under speech) and master to
-14 LUFS, -1 dBTP before the mux.
Music sourcing. Search online libraries in this order: Pixabay Music, Freesound (filter licence CC0), Mixkit,
Uppbeat. Pick by mood, tempo and a clear ending; set the score's key, bpm and chords to the track so the effects
are tuned to it, and retime the picture to its bars.
Music licence rule.
- Prefer commercial use without attribution (Pixabay Content License, CC0, Mixkit free licence).
- CC-BY only with the credit in the post caption.
- Never NC (non-commercial) or ND (no derivatives).
- Never in-app platform music (Instagram, TikTok libraries): business accounts can't use most of it and the publishing API can't add it.
- Record the track URL, author and licence in the project README; keep the file out of git.
The craft law
- One message, written as a claim. Every beat's job traces back to it (the feeling counts); anything else is cut.
- Story before polish. The script comes from
scriptwriting: a chosen structure, an ABT at its core, pain and feeling first with features as proof. Persuasion films land the message by beat 2; story films land the stakes by beat 2 and the message at the turn. Never paraphrase the source in its own order. - Never front-load. At t=0 show only what is being said; reveal each piece when the voice names it. A frame that fills instantly and then freezes is a slide, not a shot.
- Rhythm is declared, not discovered. Name the pattern (fast-fast-SLOW-hit-hold) before building. Allocate deliberate holds and a 0.3 to 0.75 s pause before the climax. Something new every 2 to 4 s. Make version 1 a notch slower than feels right; client notes on pace only ever say "hold longer".
- Sound-on or muted is decided per destination. Muted feeds need an on-screen text track that carries the story alone; sound-on platforms lead with voice and music. The hook always reads muted.
- Brand is sacred, layout is free. Quote tokens verbatim from the product's design source; scale them for video (headlines 64 to 120 px at 1080p, borders 2 to 4 px). Always set the brand font; the default font is a fingerprint.
- Real product only. Capture or faithfully rebuild real screens. Never invent UI, features, numbers, testimonials or
results; label illustrations; every on-screen number has a source in
facts.md. Generated imagery is never evidence. - Motion has a cause and a direction. Entrances ease out, exits ease in and run about 75% as long. One dominant direction per film; other vectors are reserved for meaning. Cuts land mid-motion with matched axis, direction and speed. Springs need a damping ratio of 0.7 or more unless the brief is explicitly playful (Remotion's default spring is 0.5, so always pass a config).
- Fill time with story, not wobble. No idle breathing or floating on content. An under-filled shot needs a staged reveal, a camera move with intent, or less time.
- One timeline drives everything. Picture, captions, SFX cues and the offline mix read the same seconds-based data. With a voice, the voice is the clock; otherwise the music is.
- Music has a drop; put the key visual on it. Found by measuring energy, never by trusting an auto beat grid.
- Sound is half the film. Every visible state change that should sound gets one, placed by the cinetic score on the frame the picture computes; the bed ducks under voice; the score masters to the platform target.
- The judge is never the builder. Critics are separate read-only sub-agents, and they look once.
- Trust measurements over opinions. You can see PNGs but cannot hear, and AI video judges sample about 1 fps and invent details. A critic confirms a motion claim on a frame strip before it reports it.
- Frame 0 is the thumbnail. Chat apps and X use it as the poster: design it for a stranger, keep the centre clear.
- Every format is re-laid out, never cropped. One timeline, a layout function per format.
- Tell the truth about the making. Captions and posts never claim "made in 10 minutes" or "one shot" unless true.
- No AI mannerisms. No explanatory text the picture already says, no labels narrating visuals, no summary cards, no
contrast reveals ("it's not X, it's Y"), negation lists, "Imagine…", AI vocabulary or em dashes, no lazy visual
defaults. The anti-AI pass (
qa.md§7) runs on the script while writing it and inside the Taste critic. It is never skipped.
Where we deviate from the official Remotion defaults
- Rendering. The official router says not to render unless asked. Here the deliverable is a checked file, so rendering stills, strips and drafts is always fine; the full master is rendered after the review.
- Timing source. Official markup puts timing inline on JSX nodes for Studio editing. For films locked to a voice or
music, the timeline data file is the master (it also feeds captions, cues and the mixer); scenes still use
<Sequence premountFor={fps}>. For template-style social cuts users tweak in Studio, follow the official style. - Transitions.
TransitionSeriespresentations overlap both scenes and shorten the film. Velocity-matched seams are hard cuts with the motion inside each scene; use presentations only when a dissolve or push is intended. - Audio. Remotion does no loudness normalisation or ducking. The soundtrack is premixed offline by cinetic's
score.pyand muxed after the render; in-component volume curves are for drafts. - SFX library. The official
remotion.mediasound list is mostly meme sounds. Never use those in brand films. - Studio. One Studio per project, owned by the orchestrating agent. Sub-agents verify with stills and strips.
Tools bundled with this skill
Paths: scripts live in this skill's scripts/ folder, under the base directory shown when the skill loads (~/.claude/skills/video-director/ for a standard install; a plugin install puts it elsewhere). The other skills of this suite sit next to it as sibling folders. Paths written as ~/.claude/skills/... below assume the standard install.
uv or bash scripts with --help, run from the video project root, writing under out/. Path:
~/.claude/skills/video-director/scripts/; full flags in scripts/README.md. Voice, image and sound scripts live in their
own skills.
| Need | Script |
| Contact sheet of chosen seconds | stills.py |
| Every frame of a fast move | strip.sh |
| Review sheets from a rendered MP4 (contact, phone, around t, beats) | sheets.sh |
| Automatic scan: pops, flashes, freezes, black frames, spec | qa_video.py |
| Chat-thumbnail simulation; burn a poster into frame 0 | poster_check.py |
| Render every composition and variant with master flags and measure spec and loudness | render_masters.sh |
| Platform exports (YouTube, X, LinkedIn, web, GIF) | export_platforms.sh |
| Audio/video model judgements: VO takes, music fit, muted restate | media_judge.py |
Templates in assets/: project-template/ (a runnable Remotion project with the timeline, cue, caption, motion and format
system), CREATIVE.template.md (brief to decision log), BUILDER_BRIEF.template.md and CRITIC_BRIEF.template.md.
Start a new film by copying assets/project-template/ into the product repo as video/ (its own pnpm package, excluded
from the host app's tsconfig and lint), then fill CREATIVE.md from the template. Many of the measured rules come from
shipping the Repitis launch film (a 73 s narrated promo for a children's reading app) and its 1,372-take card audio.
Working with sub-agents
- Build the shared core first (timeline, tokens, motion lib, format helper, assembly with placeholder scenes), probe-render
it on one sheet, then fan out. Up to about six short scenes, building inline is faster; beyond that, give each builder two
or three scenes. Each builder owns its scene files only and gets
BUILDER_BRIEF.template.mdfilled in. - Voice, images and sound can each be their own agent from the start, running their skill in parallel with picture.
- Critics get
CRITIC_BRIEF.template.md, one per lens. They run once, in parallel; the fixes go in and the film is done. - One heavy render at a time on the machine; parallel renders just thrash.
Hand-off
End with: the file paths (revealed with open -R on macOS), duration and formats, a one-paragraph true caption, what was
real and what was generated, music and asset licences, open issues, and where the decision log lives. No preamble, no
self-praise.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

