learn-movie

TEACH Horus by analyzing actual movies. Extracts lighting, composition, emotional cues, and director style from video files.

Features

• Scene Detection: Automatically detects scene cuts using ffmpeg.
• VLM Analysis: Uses Vision-Language Models (via scillm) to analyze keyframes.
• Memory Integration: Stores learned patterns into the horus-movies collection.

Usage

# Analyze a video file
/learn-movie analyze path/to/movie.mp4

# Analyze with specific director context
/learn-movie analyze path/to/movie.mp4 --director "Ridley Scott"

Setup

Requires ffmpeg installed. Uses scillm skill for VLM inference (e.g., pixtral-12b).

Memory + Taxonomy Integration

Follows the canonical create-context/memory_integration.py pattern with graceful degradation.

Pre-hook: recall_prior_analyses(movie_name, k=5)

• Recalls prior cinematographic analyses for the same movie before analysis begins
• Surfaces previously learned techniques to avoid redundant storage
• Enables cumulative learning across sessions

Post-hook: learn_analysis(movie_name, techniques, scenes, insights)

• Stores analysis snapshot (movie, date, technique counts, bridge tags)
• Stores individual techniques (shot type, lighting, composition)
• Stores scene-level insights for cross-movie pattern matching
• Tags: ["learn_movie", movie_name] + bridges

Bridge Keywords (keyword fallback)

Graceful Degradation

• Uses try/except ImportError at every hook callsite
• Falls back to keyword-based bridge extraction when taxonomy module unavailable
• All hooks are no-ops when common.memory_client is not importable

File Structure