What problem does it solve?
Screen recordings of chat histories, social media, and web pages used as legal evidence often contain hundreds of redundant, blurry, or transitional frames, making manual review slow and error-prone. This Skill extracts high-recall keyframes locally, filters transition and loading frames, and produces traceable evidence lead indexes without uploading sensitive footage to the cloud.
Core Features & Use Cases
- Bounded high-recall frame extraction: Uses ffmpeg scene detection plus temporal clustering, SSIM/dHash deduplication, scroll-overlap merging, and optional offline RapidOCR content-delta protection to keep frames containing new amounts, IDs, or text.
- Non-destructive evidence lead ranking: Generates a privacy-preserving evidence index and contact sheets classifying frames into closed categories (parties, transactions, communications, public statements) without storing OCR raw text.
- Budgeted multimodal audit with safety gates: Prepares small audit packages (max 8 groups/24 images, or a weak-model profile) where deletions only apply when confidence, local risk signals, and coverage-frame survival checks all pass; failures roll back transactionally.
- Use Case: A lawyer has a 3.5-minute screen recording of a WeChat conversation. The Skill reduces 411 candidate frames to about 54 traceable evidence screenshots with SHA256 hashes, then produces a curated subset via visual audit without ever modifying the base frames.
Quick Start
Extract key evidence frames from my screen recording chat-log.mp4 and generate a deduplicated, traceable screenshot set with an evidence lead index.
