video-creation

Video Creation Capability Pack

Cross-agent portable judgment for AI-produced video. Covers storytelling, motion design, audio, tool selection, and quality. CONSUMES: Brand/design artifacts (optional). PRODUCES: Professional-quality video compositions.

Step 0: Pack Prerequisites

This pack requires:

• FFmpeg — encoding and audio mixing
• Node.js ≥22 — HyperFrames or Remotion runtime
• HyperFrames CLI (npx hyperframes, pinned v0.6.97 / 2026-06-14) OR Remotion (npx remotion, pinned v4.0.477 / 2026-06-14)
• fal.ai API key (FAL_KEY) — for Seedance 2.0 video generation (optional, only if using AI asset generation)
• Codex CLI — for gpt-image-2 image generation (optional, only if using AI asset generation)
• ElevenLabs API key (ELEVENLABS_API_KEY) — for TTS, voice cloning, and AI SFX (optional)
• Fish Audio API key (FISH_API_KEY) — for cross-lingual TTS and voice cloning (optional, alternative to ElevenLabs)

Verify: ffmpeg -version && node --version && npx hyperframes --version Or run the deterministic preflight: bash scripts/verify-prereqs.sh (exit 0 = ready; 1 = ffmpeg missing; 2 = node < v22; 3 = composition CLI missing). Add --remotion to check Remotion instead.

Step 1: Context Detection

Detect the user's request type and load the appropriate reference file(s).

Multi-signal: Load all matched references. Cross-reference sections are linked within files.

⚠️ Tier-0 first for any montage / beat-sync / "photos → video" request: before picking a tool, classify the weight of the path (casual one-off vs production pipeline) per references/beat-sync-montage.md §Tier-0. Default casual briefs (≤8 assets, one-off, no brand/determinism need) to Tier 1 (CapCut/FFmpeg) — do NOT auto-route to HyperFrames/Remotion. Escalate to Tier 3 (composition framework) ONLY when a production signal (brand system, reuse, determinism, batch) is present. This prevents the over-engineering that loses casual briefs.

Step 2: Apply Rules

Read the matched reference file(s) and apply the rules directly. Rules are concrete parameters — not guidelines.

1. Detect — identify signal from user request
2. Load — read matched reference(s)
3. Apply — use the concrete rules (timing values, GSAP curves, BPM ranges, failure checklists)
4. Produce — structured findings report (see Output Format below)

Quick Rule Index

One-line summary per rule with reference pointer. Do not inline rules here — load the reference.

Storytelling (references/storytelling.md)

• 3-5s Attention Rule: Meaningful visual change every 3-5 seconds → §Pacing Rules
• Text-Shot Duration Formula: 0 words=1.5-2s, 1-3=2-3s, 4-10=3-4s, 11-20=4-6s, 21-35=6-8s → §Text-Driven Shot Duration Formula
• 50% Reading Rule: Last element entrance finishes at 50% of scene duration → §Pacing Rules
• 5-Second Scene Ceiling: Hard max (exceptions: counter, hero hold) → §Pacing Rules
• 95% Hard Cut Rule: Only 2-3 shader transitions per 6-8 scene video → §Pacing Rules
• Video Type Patterns: Product Demo / Social Short / Tutorial timing templates → §Video Type Pacing Patterns

Visual Design (references/visual-design.md)

• Easing-by-Emotion: 6 GSAP curves mapped to emotion (power2.out → smooth, etc.) → §GSAP Easing-by-Emotion Table
• 3-Ease Minimum: At least 3 different easing curves per scene → §Motion Rules
• Entrance Offset: Never start at 0.0s — offset 0.1-0.3s into scene → §Motion Rules
• Transition Duration: Min 0.3s, sweet spot 0.5s → §Motion Rules
• No Exit Rule: Never exit-animate except final scene → §Motion Rules
• Anti-Patterns: JPEG-with-progress-bar, banned effects, loop limits → §Anti-Patterns

Audio Design (references/audio-design.md)

• BPM-to-Video-Type: 5 types × BPM range × instrumentation → §BPM-to-Video-Type Mapping
• Volume Mix: Voiceover=100%, background music=10-20% → §Volume Rules
• No Vocals Rule: Explainer/tutorial music must avoid vocals and voice-like instruments → §BPM-to-Video-Type Mapping
• SFX Pre-Lead: Whoosh starts 10-20ms before visual transition → §SFX Timing Rules

Tool Selection (references/tool-selection.md)

• HyperFrames-first: HTML-native, no build, AI-friendly — default choice → §Decision Tree
• Remotion-when: React components/state required → §Decision Tree
• FFmpeg-direct: Processing/encoding only, no composition → §Decision Tree

Production (references/production.md)

• 17 Agent Failure Modes: Timing, animation, composition errors → §Agent Failure Modes Checklist (17 Items)
• 5 Prevention Patterns: Skill loading, DESIGN.md, skeletons, CLI loop, chaining → §Prevention Patterns (5 Items)
• Render Pipeline: scaffold → compose → preview → validate → render → export → §Render Pipeline

Quality (references/quality.md)

• Export Settings: Per-platform (YouTube/TikTok/Instagram/Twitter) → §Platform Export Specifications
• Upload codec Rule: H.264 High + AAC MP4 faststart for ALL platforms; AV1 = delivery only, NOT an upload codec → §Codec & Quality
• WCAG Accessibility: ≥99% caption accuracy, 4.5:1 contrast, WebVTT → §Accessibility (WCAG)
• CRF master vs web: CRF 18 slow = upload master; CRF 23 = web cut (+6 CRF ≈ half file size) → §Export Settings
• Loudness Rule: no single cross-platform LUFS; YouTube -14 documented, TikTok/IG unverified; -16 LUFS pragmatic master → §Loudness Normalization

AI Asset Generation (references/ai-asset-generation.md)

• Seedance Default Rule: Video clips → Seedance 2.0; 4K needed → Kling 3.0; existing Runway → Runway Gen-4 → §Decision Tree
• Multi-Shot Planner Rule: 4K + multi-shot → Kling 3.0 AI-Director (≤6 shots/15s native); else Seedance manual "Shot N:" → §Multi-Shot: Kling AI-Director vs Seedance
• gpt-image Model Guard: Pin gpt-image-2 for ALL work; gpt-image-1.5/1-mini/chatgpt-image-latest all shut down 2026-12-01, DALL·E shut down 2026-05-12 — do NOT route cheap drafts to 1-mini, use gpt-image-2 quality="low" → §Model Lineup
• Endpoint Selection Rule: text-only → text-to-video; have image → image-to-video; multi-ref → reference-to-video → §Seedance Endpoint Selection
• Submit-Then-Poll Rule: Never subscribe(), always submit-then-poll with 5s/10s/120s schedule → §Async API Pattern
• Tiered Generation Rule: Draft 480p/Fast → approval → Final 1080p/Standard → §Cost Control
• Request Hashing Rule: hash(model+prompt+settings) before every API call, re-roll uses attempt_number → §Request Hashing
• Prompt Consistency Rule: gpt-image-2 invariant anchoring + Seedance @character:<id> → §Visual Consistency Rules
• Path Split Rule: Remotion assets → public/generated-{images,clips}/; HyperFrames assets → assets/generated-{images,clips}/ → §File Path Convention
• TTS Tool Selection: English expressiveness → ElevenLabs v3; CJK/cross-lingual → Fish Audio S2 Pro; simple/cheap → OpenAI tts-1-hd → §TTS Voiceover Rules
• Voice-First Timing Rule: Generate TTS voiceover BEFORE composing video scenes — voiceover duration drives scene timing → §Voice Pipeline Integration
• Clone Minimum: Fish Audio 10–15s sample; ElevenLabs 30–60s (IVC) or 30min (PVC) → §Voice Cloning Rules
• SFX Source Rule: Diegetic/scene-tied → Seedance native audio; specific/imaginative/looping → ElevenLabs SFX API → §AI Sound Effects Rules

Validation Scripts (scripts/)

• failure-mode-precheck.sh: Deterministic linter — greps a composition for the 6 banned timeline anti-patterns (Date.now/Math.random/setInterval, repeat:-1, async-await, visibility, inline opacity:0); exits 1 on hit → scripts/failure-mode-precheck.sh <file|dir>
• verify-prereqs.sh: Step-0 preflight — ffmpeg + node≥22 + HyperFrames v0.6.97 (or --remotion v4.0.477) with explicit exit codes 1/2/3 → scripts/verify-prereqs.sh

Beat-Sync Montage & Complexity Routing (references/beat-sync-montage.md)

• Tier-0 Weight Rule: classify casual one-off vs production BEFORE tool choice; default casual to Tier 1, never auto-Tier-3 → §Tier-0 Decision
• CapCut Beat Sync: auto beat-markers on waveform, free tier, mobile+desktop (not Web) — Tier 1 default for casual → §Tier 1
• FFmpeg beat-sync recipe: zoompan Ken Burns (8000×4000 prescale = PARTIAL jitter fix, never-merged u

Content truncated.