foryourhealth111-pixel/Vibe-Skills

What it solves

VibeSkills is a general-purpose orchestration layer for AI agents that solves the problem of "skill fatigue" and inefficient tool use. When agents have access to a large library of local specialized skills (tools), they often struggle to discover the right ones, repeatedly invoke ineffective tools, or fail to plan complex multi-step workflows. VibeSkills reduces the cognitive burden on the user and the agent by automatically routing and orchestrating these local skills into a verifiable delivery process.

How it works

VibeSkills implements a structured five-stage process to move a task from request to delivery:

1. Requirement Confirmation: Clarifies goals, constraints, and expected deliverables before work begins.
2. Level Recommendation: Suggests a workflow level (L for manageable multi-step work or XL for larger, potentially parallel tasks) based on scope and dependencies.
3. Skill Organization: Scans local skill folders, reads SKILL.md files to shortlist candidates, and assigns specific skills to task modules with defined deliverables and checks.
4. Execution and Recording: The agent executes the plan, utilizing TDD (Test-Driven Development) for code tasks and recording the state of each work unit.
5. Result Verification: Compares actual outputs against the planned items to ensure all requirements are met before final acceptance.

Who it’s for

It is designed for users and developers who maintain large libraries of reusable AI skills/tools and want an agentic framework that can autonomously plan, select, and execute complex professional tasks across diverse domains without manual micro-management.

Highlights

• Efficient Resource Use: Benchmarked on SkillsBench, it increased average task scores by 21.12% while reducing token consumption by 29.6% and tool calls by 33.1%.
• Intelligent Routing: Instead of passive triggering, it decomposes tasks first and selects only the most relevant skills from a large pool (e.g., 100+ skills).
• Verifiable Workflows: Integrates a rigorous checking phase (e.g., 17 checks for a single ML experiment) to ensure cross-deliverable consistency.
• Scalable Library: Supports the addition of personal, team, or third-party skills without increasing token overhead for every task, as it uses local discovery and indexing.