windagszip

You can compress SKILL.md files by using embedding-based redundancy detection to remove duplicate chunks while preserving behavior. This reduces token usage and context window consumption across skills in the WinDAGs repository.

Embedding-based deduplication analyzes SKILL.md chunks using vector embeddings to identify semantically redundant text clusters. It enables targeted removal of duplicate content to reduce token usage while maintaining original functional behavior.

Generate ablation variants by applying cluster-aware variant generation to compressed SKILL.md content. This produces multiple compressed options that support graded-eval workflows to map token savings directly to behavioral performance.

Yes, you can evaluate behavioral integrity after markdown compression using optional graded-eval workflows. These workflows map the token savings achieved through deduplication directly to the preserved skill performance and quality metrics.

Yes, windagszip supports compressing single or multiple skills in the WinDAGs repository. It applies embedding-based redundancy detection across multiple SKILL.md files to reduce overall context window usage simultaneously.

A limitation of embedding-based skill compression is the potential loss of behavioral nuance if redundant chunks contain context-critical details. Graded-eval workflows are recommended to verify that token savings do not degrade performance.