What's this about?
As agents store more of their procedural knowledge in reusable skill packages loaded at inference time, the libraries grow — and the problem becomes exposing the smallest sufficient executable context within a limited budget. SkillZip (73 upvotes) compresses the skill graph while preserving each skill's contract.
What exactly is a skill's "contract"?
A contract is the promise a component makes about its interface: what it needs as input, what it guarantees as output, what it depends on. Compression that preserves the contract means you can shrink the representation — merging overlapping skills, dropping redundant description — without any caller noticing a behavioral difference. It's the same discipline as refactoring: change the internals freely, never break the promise.
Why is this worth noticing?
- It's a research response to a problem that only appeared once skills became widespread, and it lands the same day the community is actively publishing and curating skill collections.
- Context is a hard budget. Every token spent describing available skills is a token unavailable for the actual work, so compressing the skill library directly buys capability at no model cost.