Vibe-Skills

VibeSkills addresses a problem that appears once you install many agent Skills: the agent picks the wrong ones, skips steps or loops. You invoke it as a Skill (`$vibe`, `/vibe` or the host's syntax) in any agent that supports local Skills.

It first clarifies the goal and delivery requirements, splits the task into verifiable work units, then shortlists candidates from a local index of Skill names and descriptions and reads only the retained `SKILL.md` files. Selected Skills are assigned to specific parts of the work, and a final acceptance check runs across all outputs. Plans, execution status and acceptance reports are saved so a later session can resume and a review can compare plan with result.

The authors report that on an adapted SkillsBench setup with all 195 Skills installed, using DeepSeekV4Flash-VE with OpenHands, version 4.1.0 raised mean verifier reward by 21.12 points while cutting total tokens by 29.6% and tool calls by 33.1%. The README notes that planning still adds context cost that grows with candidate count and task complexity.