The prompt method behind the AI short film PJ Ace couldn't stop tweeting about

What it does It turns a rough idea into a structured, model-ready video prompt for Sora, Kling, Veo, or Seedance. The core is a 5-stage structure — theme, atmosphere, camera, timed beats, ending — plus 21 genre templates and a library of 50 camera moves. There’s also a no-install browser-based prompt builder if you don’t use Claude Code.

The interesting bit The whole thing is reverse-engineered from Zombie Scavenger, a 3-minute AI short by a Chinese creator called Mx-Shell that a Hollywood director publicly called one of the best short films he’d seen in years. Mx-Shell gave away his full method on a livestream; this repo digests ~130,000 characters of his materials into a reusable system. The trick that carries the method: naming real camera bodies and lenses (“simulated IMAX + Panavision C-series, 35mm, f/4”) instead of vague keywords like “cinematic, 4K, epic.”

Key highlights

• 21 IP-stripped genre templates, from claymation to found-footage horror to food ASMR
• Mx-Shell’s original prompts preserved (Chinese body, English headers), including the full Zombie Scavenger prompt
• Failure→fix gallery of common bad outputs and their corrections
• Per-model negative-prompt prefabs and a camera-movement SOP for 5 film genres
• Bilingual throughout (English + Chinese), with a one-page cheat sheet

Caveats

• The README is heavily cross-linked to the author’s own courses and web tools; the repo is partly a funnel for those
• Original prompt bodies stay in Chinese — English readers get the generalized templates, not Mx-Shell’s verbatim voice
• “Eval-tested” is claimed but the evaluation methodology isn’t shown in the README

Verdict Worth a look if you’re prompting text-to-video models and getting generic CG sludge back — the camera-language approach alone is a genuine upgrade. Skip it if you wanted a model, a pipeline, or anything that runs without an LLM in the loop.