Video Production AI

What's in this pack

This pack collects the seven video production assets that show up in every team shipping daily AI-generated short-form content (TikTok / YouTube Shorts / Instagram Reels). Three are headline frameworks; four are the headless pipeline templates that wire them into a working assembly line.

Why this pack matters

Every "AI video" tool you've seen on Twitter is a UI wrapper — fine for one-offs, useless for shipping 10 videos a day. Real production pipelines are headless: you write code, render in CI, upload via API, repeat. This pack collects the seven pieces that compose into that pipeline. None of them require a designer in the loop.

The three headline frameworks each take a different bet:

• Remotion is the "video is code" bet. You write React components that get rendered into MP4 frames by a headless Chromium. You get version control, hot reload, programmatic data binding, and a templating language any frontend dev already knows. Best for branded series where every video has the same shell with different data.
• OpenMontage is the "LLM as editor" bet. Give it a script, it picks b-roll from a stock library, sequences cuts, and renders. Best when you want zero-touch automation and don't care if every cut is perfect.
• MoviePy is the "Python all the way down" bet. A direct ffmpeg wrapper with a clean API for cutting, concatenating, overlays, and effects. Best when you want full control and don't need React.

The four pipeline templates are the difference between "I can render one video" and "I render 50 a day on a cron." They cover the actual operational steps: scripting, captions, b-roll, and multi-platform publishing.

Install in one command

# Install the entire pack
tokrepo install pack/video-production-ai

# Or pick the framework you actually need
tokrepo install remotion
tokrepo install openmontage
tokrepo install moviepy

The TokRepo CLI installs the framework adapters into your AI tool — Claude Code subagents into .claude/agents/, Cursor rules into .cursor/rules/, AGENTS.md entries for Codex CLI. The Node/Python install is your job (npm install / pip install). TokRepo wires the prompts so an agent like Claude Code knows when to invoke Remotion vs MoviePy.

Common pitfalls

• Pre-fetch all assets before render. A render that fetches a stock clip mid-pipeline blocks for 30+ seconds and frequently times out. The b-roll prefetch template (asset #6) downloads everything to a local cache first; never skip it on production runs.
• Don't render at 1080p when you publish 9:16. TikTok / Shorts / Reels all want 1080×1920 vertical. Rendering 1920×1080 then cropping wastes 50% of your render time. Set canvas dimensions explicitly in Remotion's <Composition> and MoviePy's resize().
• Captions misalign by 0.3s on long takes. Whisper's word timestamps drift on continuous audio. Force-align with WhisperX or pad the SRT entries by 100ms; viewers notice misaligned captions before any other quality issue.
• Music drowns out narration without sidechain. A ducking compressor on the music track keyed to the VO is 90% of the perceived quality jump. Both MoviePy and ffmpeg can do this — see the audio post-processing recipe in the pipeline template.
• YouTube Shorts has hidden length constraints. Up to 180s as of 2026, but the algorithm still favors <60s. The multi-platform publisher template caps at 58s by default for Shorts and 60s for TikTok/Reels.

When this pack alone isn't enough

This pack covers production pipelines, not creative direction. If your videos look bad, the issue is usually:

• No real source content. AI-generated b-roll, AI-generated voiceover, AI-generated captions add up to slop. Pair this pack with the Browser Automation pack to scrape real source data, the Voice AI Stack pack for production-grade TTS (ElevenLabs, Cartesia), and an actual editorial point of view.
• One template forever. A pipeline that ships 10 videos a day with the same shell will get downranked within 2 weeks. Build at least 3-4 visual templates and rotate.
• No human in QC. Even with the best pipeline, ~5% of generated videos will have a glaring error (wrong logo, mistimed caption, blank frame). A human spot-check before publish keeps your channel alive.

The pack itself is the assembly line. The script, the data, and the editorial taste are still yours to bring.