Claude Skills v2 has been released, representing a major upgrade to how skills work in the Claude platform. The centerpiece is a new skill creation skill that employs multiple sub-agents to test, benchmark, and improve skills automatically. The workflow begins with Claude interviewing the user about their intent for creating a skill, followed by research into relevant documentation. It then generates the skill.md file and necessary assets, along with comprehensive test cases and evaluations. The system spawns two parallel, isolated sub-agents - one using the created skill and one without it - to compare performance. Both agents' outputs are tracked for token consumption and time to completion. Results are aggregated and graded against the initially created test cases, with outputs displayed in an HTML viewer in the browser for user review. The system then accepts user feedback and iterates on the skill until no further improvements are needed. This new version can be used both for creating new skills and upgrading existing ones.
Claude Skills v2 centers around a new skill creation skill that uses multiple sub-agents to test, benchmark, and improve skills
High confidence
The skill creator interviews users about their intent and researches specific documentation
High confidence
The system spawns two parallel sub-agents that are completely isolated from each other
High confidence
One sub-agent uses the created skill while the other does not, allowing for comparison
High confidence
The system tracks token consumption and time to completion for performance metrics
High confidence
Results are graded against test cases and displayed in an HTML viewer in the browser
High confidence
The system iterates on skills based on user feedback until complete
High confidence
The new skill creator can upgrade old skills in addition to creating new ones
High confidence
The creator's overall position toward the main topic discussed.