ai-filmlisted

Produce multi-shot narrative short films (60–240s) with AI video generation — story architecture, shot design, locking a real person's face across dozens of shots, dialogue, transitions, black-and-white and period looks, subtitles timed from audio, and multi-language/multi-ending variants. Use when asked to make a short film, movie, remake, cinematic video, trailer, or any multi-scene narrative video with recurring characters. Also use when an existing AI video looks "not cinematic", when a viewer says they can't follow the story, when characters drift or get swapped between shots, when someone's face stops looking like them, when props or wardrobe change between scenes, when a transition swallows a line, when "black and white" comes back in colour, or when a finished video won't play in a browser.
KeWang0622/ai-film · ★ 1 · AI & Automation · score 72

Install: claude install-skill KeWang0622/ai-film

# AI Film Production Build a narrative short from a premise. Two stages: **generation** (paid, slow) and **post-production** (free, fast, endlessly re-runnable). Keeping them in separate scripts is the most important structural decision here — transitions, grading, loudness, subtitles and delivery get re-run dozens of times without spending again. ``` shots_<project>.py scenes: shot design, identity, props, audio rules, dialogue check_script.py lints the script; --script prints the dialogue alone free generate.py → out_<project>/*.mp4 via pika_client.py (stdlib only) paid audit_faces.py → audit_<project>/*.png face contact sheets free finish.py → concat → grade → (subtitles) → (score) → cut free verify_film.py measures the cut: runtime, colour, music leak, banding free deliver.py → MASTER / full / share, two-pass linear loudness free ``` Copy `templates/` into a working directory, then `cp shots_template.py shots_<project>.py` and edit that. Every other script is generic and routes by `--project <name>`. Run `check_script.py` before spending anything. --- ## The one rule everything else follows from **The model executes descriptions. It does not execute negations. And it fills every space you leave it.** Nine times out of ten a defect is fixed by *changing what you described*, not by adding another "NO …". Measured, repeatedly: | Defect | Adding a negative | Changing the