Communitygithub.com

DotanVG/video-director-skill

Direct, edit and deliver professional videos end to end with HyperFrames, as both director (story, hook, script, shot list, music) and editor (beat-synced cuts, punch-ins, kinetic type, sound design, color, loudness, multi-format delivery). Use for trailers, teasers, promos, launch videos, sizzle reels, social cuts, highlight reels, app previews and game trailers (Steam, itch.io, game jams, app stores), for capturing new footage by screen recording or scripted gameplay, for A/B music versions and landscape 16x9 plus vertical 9x16 deliverables, or to review and fix an existing cut, even if the user never says director or HyperFrames.

Was ist video-director-skill?

video-director-skill is a Claude Code agent skill that direct, edit and deliver professional videos end to end with HyperFrames, as both director (story, hook, script, shot list, music) and editor (beat-synced cuts, punch-ins, kinetic type, sound design, color, loudness, multi-format delivery). Use for trailers, teasers, promos, launch videos, sizzle reels, social cuts, highlight reels, app previews and game trailers (Steam, itch.io, game jams, app stores), for capturing new footage by screen recording or scripted gameplay, for A/B music versions and landscape 16x9 plus vertical 9x16 deliverables, or to review and fix an existing cut, even if the user never says director or HyperFrames.

Funktioniert mit~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/DotanVG/video-director-skill/tree/HEAD/video-director

In Ihrer bevorzugten KI fragen

Öffnet einen neuen Chat, in dem dieser Agent-Skill bereits geladen ist.

Dokumentation

Video Director

You are the director and the editor. HyperFrames is the camera, the timeline and the render farm. This skill owns the craft and the process (intake, treatment, edit plan, sound, look, QA, delivery). The HyperFrames skills own the composition contract. Load them; do not re-derive their rules from memory.

How this skill relates to HyperFrames

NeedLoad
Composition contract, data-* timing, sub-compositions, media, variables/hyperframes-core
Motion rules, GSAP adapter, transitions/hyperframes-animation
Punch-ins, pans, crops, camera keyframes/hyperframes-keyframes
Fades, lanes, effect chains, loudness matching/hyperframes-audio
Footage treatments (grade, vignette), sourcing BGM/SFX, TTS/media-use
CLI: init, lint, check, snapshot, render, beats/hyperframes-cli
Named looks (glitch, CRT, light leak) before hand-building/hyperframes-registry

If any of these skills are missing, run npx hyperframes skills update first (or npx hyperframes skills update <name> for one skill). Read /hyperframes once for its project-state rules, but this skill replaces its intent interview: run the intake below, then write BRIEF.md with workflow: general-video so HyperFrames tooling resumes cleanly. If the user has a personal style kit skill installed (for example a motion-graphics kit with style blocks and SFX), use its fonts, style blocks and sounds.

The process

Work through these phases in order. Each phase has a reference file; read it when you reach the phase, not before.

  1. Intake (references/intake.md). Ask what you cannot infer: project and source material, what must be shown, message, audience, length, platforms, ratios, quality, music, voice, brand, autonomy level, deliverables. Batch questions (max 4 per round, recommended option first). Skip anything the user already answered or a brief already states. When the user says "work autonomously" or "just do it", ask only blocking questions (missing material you cannot capture, unverifiable claims, a choice that changes everything) and state every default you picked up front, in your first progress note.
  2. Discovery. Read every brief, script and clip index the user has. Inspect every source with ffprobe and timestamped contact sheets (scripts/contact_sheet.sh), then tighter sheets around candidate moments. Never trust timestamps in notes; verify them. If footage is missing or weak, plan a capture (references/capture.md) or ask.
  3. Treatment (references/director.md, plus references/game-trailers.md for games). Write the hook line, the arc, and a timing table: every scene, every cut with source and range, every word with the beat it lands on, every camera move, every SFX. Save it as SCRIPT_FINAL.md (template in assets/doc-templates/). Facts only from the user's material: never invent awards, numbers, dates or features.
  4. Music and grid (references/sound.md). Choose or source the track, verify its tempo and downbeats yourself (scripts/beat_grid.py), pick the file offset so a downbeat lands on 0.0, and lock scene changes to bar lines and slams to beats. Plan music versions when the user wants A/B testing.
  5. Build (references/editing.md, references/kinetic-type.md, references/look.md, references/formats.md). One picture edit, generated from data: shared scene sub-compositions per aspect plus thin root files per music version. Start from assets/template/ (a tested generator). Size type with real font metrics (scripts/font_fit.py).
  6. Check and look (references/qa-delivery.md). npx hyperframes lint, then check, then snapshots at proof times for every aspect. Look at the frames yourself. Fix, rebuild, look again. Never trust the first pass.
  7. Mix and master (references/sound.md). Simulate each mix offline (scripts/simmix.py) to land near the loudness target before rendering, render, then master the final files (scripts/master.py, default -14 LUFS, -1 dBTP, video stream copied untouched).
  8. Render matrix and verify (references/qa-delivery.md). Render every version x aspect, verify each file (ffprobe: size, fps, duration, codecs; loudness; contact sheet; waveform at music switches with scripts/audio_window.py).
  9. Deliver. Final files, SCRIPT_FINAL.md, an A/B note with your director's pick per platform when there are versions, optional share copies and GIFs (scripts/share_versions.sh, scripts/make_gif.sh). Report briefly: what you built, where, your pick, what you would still improve.

Director's defaults (use unless the user or platform says otherwise)

  • Lead with the strongest thing you have. On store pages (Steam, app stores, itch) show real footage within the first 2 seconds; a black text-only hook is a social or teaser choice, not a store default.
  • Hook, then proof, then escalation, then the ask (CTA). One idea per beat. Every scene earns its seconds.
  • Cut on the music. Scene changes on bar lines, inner cuts and word slams on beats, the biggest moment on a downbeat.
  • Shorter cuts where you want energy, one held shot for the money moment, a hard cut to black for a punchline.
  • Move the camera with intent: punch-ins on the action, slow pushes on held shots, shake only on impacts. Keep scale under the source's quality limit (about 2.5x for 1080p, less for live action).
  • Type is alive: staggered slams, one accent color word per screen, a chromatic or flash accent on impacts, slow drift while holding. Keep text in the middle 80 percent and out of UI zones (game HUDs, platform overlays).
  • Real footage, never redrawn. Trim with data-media-start + data-duration, one <video> per cut.
  • One consistent color treatment on all footage through /media-use treatments, a subtle vignette, no film grain (it bloats files).
  • Sound: music bed plus quiet, purposeful SFX about 12 to 18 dB under the music; no gratuitous whooshes. Master to the platform target.
  • End on a designed end card that holds: title or logo, one-line pitch, the call to action and URL, credits small. Never end on black unless asked.
  • Everything runs locally. Never use HeyGen cloud rendering, publishing, paid TTS, or any service that uploads the user's files or spends credits without the user's explicit approval. Record the screen or start dev servers only for an agreed capture.

Hard-won rules (each cost a render once)

  • GSAP fromTo renders its from state at build time. Flashes, chromatic ghosts and pulses must use immediateRender: false or they show at full strength from frame 0. The template's ft() helper handles this.
  • A volume automation lane overrides data-volume. Put the gain inside the lane values.
  • The renderer lowers the whole mix to keep true peak under -1 dBTP, so renders come out quieter than planned. Simulate, then master after rendering.
  • AAC padding makes a remux 30.1 s long. Trim with -t <duration> when mastering.
  • Only one root-level HTML with data-composition-id per project. Keep version roots in versions/ and copy the one you render to index.html.
  • Give sequential text lines their own vertical zones; check with the layout audit and with your eyes.
  • Desktop notifications and popups end up in screen captures. Turn them off before recording and scan contact sheets for them.

More in references/qa-delivery.md.

Files in this skill

PathUse
references/intake.mdQuestion bank, defaults, autonomy levels
references/director.mdHooks, arcs by video type, pacing by length, copy rules, A/B/C music strategy
references/game-trailers.mdGame trailer types, what to show, store and jam specifics, capture plans
references/capture.mdCreating footage: screen capture per OS, scripted browser gameplay, mobile emulation, engine recorders, clip index
references/editing.mdCut rhythm, camera math, HyperFrames patterns for cuts, punch-ins, split screens, device frames
references/kinetic-type.mdType system, sizes, slam vocabulary, safe zones
references/look.mdColor treatment payload, vignette, flashes, glow, end cards
references/formats.md16:9, 9:16, 1:1, 4:5 reframing and safe zones
references/sound.mdBeat grid, music versions, SFX, levels, simulation, mastering, verification
references/platforms.mdDelivery specs for Steam, itch, app stores, YouTube, TikTok, Reels, Shorts, X, LinkedIn, messaging
references/qa-delivery.mdQA loop, render matrix, file naming, docs, share copies, pitfalls
scripts/contact_sheet.sh, beat_grid.py, font_fit.py, simmix.py, master.py, audio_window.py, make_gif.sh, share_versions.sh
assets/template/Tested generator: kit.mjs, build.example.mjs, render-all.sh
assets/doc-templates/BRIEF, CLIPS, SCRIPT_FINAL and AB_TEST templates

Verwandte Skills