Skill libraries should be compressed as executable graphs, not as text — 3.46x smaller with 99.2% of dependencies intact

arXiv 2608.05604·medium signalXBlueskyLinkedInCopy link

SkillZip identifies the mismatch in current skill systems: they retrieve skills as packages but compress them as prose, which destroys the execution contract. It performs contract-preserving compression over section-level graphs, rewriting recurring valid motifs into reversible macros while preserving boundary signatures and dependencies, achieving 3.46x compression with 99.2% dependency preservation, 98.7% verifier reachability, and up to 12.2 points of task improvement. A companion pass (ReZip) folds new skills in and revises macros from execution feedback, and the approach is demonstrated across libraries from 200 to 100,000 skills — directly relevant to anyone whose `.claude/skills` or agent playbook directory has outgrown its context budget.