classify-media-relevance

What problem does it solve?

Release parsers and crawlers frequently capture both editorial media (screenshots, demo videos, diagrams) and irrelevant site chrome (avatars, logos, tracking pixels). This Skill helps the pipeline keep meaningful media in a release's media[] array while dropping junk that would clutter the UI or waste storage, prioritizing precision over recall.

Core Features & Use Cases

• Deterministic pre-checks: low-cost rules to drop tracking domains, unsupported content-types, tiny tracking pixels, oversized files, and reuse existing uploads to avoid redundant work.
• Signal-driven classification: weighs alt text, filename patterns, URL paths, page position, dimensions hints, frequency across releases, and contextual heuristics to decide keep or drop with confidence levels.
• Structured output for pipelines: returns a JSON array preserving input order with decision, confidence, and a short reason so callers can apply conservative fallbacks for low-confidence drops.
• Use case: run during parse and crawl-mode fetches to populate a release's media[] array while skipping feed fetches that already scope media.

Quick Start

Classify the provided media items using their URL, alt text, filename, and surrounding content and return a keep or drop decision with confidence and a concise reason.