Code-built films, no video model: exploraX's Cinematic Skill, the ESCAPEMENT Blender short, and a talking-head editing skill pair

Cinematic Skill (MIT, about 200 stars since Nov 9). From exploraX (@exploraX_), who built it from what broke while making a five-minute Al Capone documentary with an agent. Ask for 'a 90 second documentary about X' and, per the README, the agent researches and fact-checks a script (one idea per line), times it with a free scratch voice (macOS say or espeak-ng) and swaps in any TTS or a real voice later, re-timing automatically. It pulls public-domain images from Wikimedia Commons, records each licence, refuses non-free files and writes credits.md for you. Then it builds the whole film as one continuous camera move over a designed 'world': cork evidence walls, maps with routes, torn clippings, stamps and highlighter swipes that land on the exact narrated word. Frames are drawn in Python with Skia, sound effects are synthesised (nothing to license), and it outputs four caption styles plus an SRT. It stops for your approval after the beat table, the look, and a 30-second draft. The author says the 92-second example renders at 1080p in about 6 minutes on a 16 GB laptop. Install with npx skills add Moh4696/cinematic-skill (Claude Code, Codex, Cursor, Gemini CLI, OpenCode). His launch post got about 435 likes, and a Persian-language repost by @Yarad_ai passed 950.

ESCAPEMENT (MIT, about 127 stars). Leonxlnx's 36-second car-and-clockwork short, made by Claude Code (Opus 5.5) driving Blender 5.2 headless from Python on a laptop with an AMD Radeon 780M integrated GPU. The repo is worth reading as a production template more than as a tool. Five builder agents worked on the five acts in parallel, a director agent critiqued each act, and the iteration logs are in notes/. Every Blender render goes through a global GPU lock, because parallel EEVEE renders crashed the laptop once. An exposure check scores every still for clipped highlights and crushed blacks. The edit, grade, halation, grain and sound mix are also code (numpy + ffmpeg), cut to the music's 120 BPM grid. Sounds are CC0 with a manifest. It pairs well with this morning's Fable 5.2 Blender films post: same idea, different model, with the whole pipeline published.

cangshu-koubo-video-skills (Apache-2.0, small but practical). Two Claude Code skills from QQTangBa for Chinese talking-head (koubo) videos on Apple silicon. The base skill cuts flubs, repeated takes and long pauses (keeping natural pauses under 0.4 s), smooths each cut with an 8-frame optical-flow blend similar to Premiere's Morph Cut, adds a subtle 'breathing' zoom and emphasis push-ins, and burns in auto-aligned captions with highlighted keywords. It renders a 9:16 video and refuses to deliver if lip-sync drift is over 20 ms, per the README. The motion skill alternates the speaker with hand-written Remotion scenes (data cards, kinetic type, matting via macOS Vision) in 3:4. Transcription is local with mlx-whisper. Docs are in Chinese, and Remotion's own licence applies.

Signals (not hot by our rule): GitHub stars plus authors' and community posts on X. None is on GitHub Trending today or picked by a tracked curator. Features and timings are the authors' claims. We haven't run them.

Our take: Cinematic Skill is the one to try first if you make explainers or history and true-crime shorts. It's deterministic, licence-aware and cheap to re-voice. ESCAPEMENT is the best public reference yet for structuring a multi-agent Blender film. The koubo skills are a good model for automating your own talking-head cutdowns, even if you don't work in Chinese.