Finderchangchang/brewreel

BrewReel is a video-generation skill suite designed for LLM coding assistants, built around a "cheap-model writes storyboard, components render the video" philosophy and originally published as promo-video-skill / 蒸馏视频. It ships three concrete integration paths: AI coding agents such as Claude Code, Codex, and opencode that can read SKILL.md install the repo as a skill and let the assistant write the storyboard, run validation, and render; a dedicated DeepSeek Harness plugin published to npm as dsh-brewreel that exposes 7 tools for validation, rendering, and inspection (still flagged as not yet tested against real DeepSeek models); and a standalone scripts/llm_make.py that calls any OpenAI-compatible endpoint, defaulting to DeepSeek, so non-agent environments can still feed a brief in and get a video out. Rendering is powered by Remotion 4.0 and produces 1080×1920 (journey defaults to 1080×1350) MP4s at 30 fps with on-the-fly synthesized soundtracks at –16 LUFS, sidestepping copyright issues by beat-mapping music to shot cuts.

The typical workflow starts with the user writing a product brief; the LLM reads SKILL.md, selects camera shots, drafts copy, and fills in a storyboard.json that only contains selectable fields plus text — no coordinates, no code. A series of prebuilt components defined under styles/ then handle layout, motion, pacing, subtitle line-breaking, and safe-area checks, while make.mjs chains validation → music → render → composition → frame inspection into a single command that refuses to deliver anything marked ✗. Built-in compliance covers 《广告法》 extreme-word rules plus industry-specific red-lines for six verticals (software, food, ecommerce, education, beauty, travel), and subtitles can be Chinese or English with an additional scan that rejects English reels which mix in stray CJK characters. The three preset recipes — cards (9:16 default feed of bordered white cards), quiz (misconception → multiple choice → reveal → glossary card), and journey (4:5 / 9:16 mascot traversing a paper-cut city) — give the same engine three distinct narrative rhythms, and example storyboards for all nine industry + recipe combinations live under examples/, styles/quiz/examples/, and styles/journey/examples/.

Dependencies are non-trivial: Node.js 18+ is required, with 20 LTS recommended and Node 22 actively tested; the DeepSeek Harness plugin path additionally requires Node 22.19+ on the 22 line or any 24+. Python 3.10+ is needed because the music synthesis scripts import numpy and scipy, although llm_make.py itself only uses the standard library. First-run downloads add up to several hundred megabytes of rendering dependencies plus roughly 110 MB of Chrome Headless Shell, and the supported operating systems are Windows x64, macOS 15 or newer, and Linux distributions with glibc ≥ 2.35 (Alpine and NixOS are explicitly unsupported because they lack libnss3, libgbm, libasound2, and related shared libraries). The project is Apache-2.0 and commercial-friendly, but redistribution must keep the LICENSE and preserve the NOTICE attribution line, and Remotion's commercial license still applies — companies with four or more employees that profit must buy a Remotion Company License separately, a constraint the Apache-2.0 grant does not override.