claude-real-video optimizes video inputs for LLMs

Local pre-processing of video files is a critical stopgap for multimodal LLMs, as directly uploading large video files is prohibitively expensive and often exceeds context limits. Scene-aware heuristics represent a major improvement over naive fixed-interval frame sampling by drastically reducing input tokens without losing content. Keeping the processing pipeline entirely local guarantees privacy and eliminates additional hosting or API pre-processing costs. Merging visual keyframes with transcribed audio provides a comprehensive multi-modal context that is easy for current models to parse. However, the dependency on local computing resources for transcription and video processing might limit throughput for bulk processing on low-spec hardware.