foryourhealth111-pixel/Vibe-Skills

VibeSkills – an orchestrator for AI‑agent Skills

What it is

• A general‑purpose “Skill” manager that sits between a language‑model‑based agent and the collection of locally‑installed Skills (reusable tool‑wrappers, scripts, or workflows).
• Its job is to plan, select, and sequence the right subset of those Skills for a given complex task, then verify the final deliverable.

Why it matters

• In many AI‑agent frameworks a user must manually tell the model which tool to call, or the model ends up flailing through a huge library of installed tools.
• VibeSkills reduces that friction by:
  1. Clarifying the task and turning it into a concrete plan.
  2. Discovering which local Skills are available (by reading their SKILL.md metadata).
  3. Short‑listing only the truly relevant Skills for each sub‑task.
  4. Orchestrating them in a dependency‑aware order (the “L” or “XL” workflow levels).
  5. Running checks after each step and a final verification before delivery.

How it works

Stage	What happens
1. Requirement confirmation	The system asks the user to approve a clear statement of goal, constraints, and expected outputs before any work starts.
2. Level recommendation	It suggests an L (single‑threaded, manageable) or XL (parallelizable, larger) workflow based on task size and dependencies.
3. Skill organization	Scans the configured Skill folders, reads each SKILL.md, and assigns the best‑fit Skills to the individual work units.
4. Execution & recording	The underlying agent runs the plan, optionally using test‑driven development for code tasks. All successes, failures, and blocked steps are logged.
5. Result checking	A set of verifiable checks (e.g., required files, data contracts, reproducibility) are run; any unmet check blocks final acceptance.

Demo case – a machine‑learning experiment

• The repository shows a full end‑to‑end run where a task to produce a reproducible classification experiment was completed.
• From 100+ installed Skills, VibeSkills selected 7, organized them into 5 work groups and 10 work units, and performed 17 verification checks. All checks passed, yielding data audit, statistical review, figures, a scientific report, and a slide deck.

Performance claim

• Benchmarked on SkillsBench (a suite that measures how well agents use Skills), VibeSkills v4.1.0 raised the mean verifier reward by +21.12 pp, cut total token usage by ‑29.6 %, and reduced tool calls by ‑33.1 % compared with a baseline that did not use VibeSkills.
• The improvement comes from better planning and selective Skill use, not from calling more tools.

Installation & quick start

• A one‑click “Install VibeSkills” button links to a short install guide (docs/install/README.en.md).
• After installation, the Quick‑start guide walks a user through confirming a requirement, picking a level, and letting VibeSkills orchestrate the work.

Who might use it

• Researchers or engineers who maintain a large library of reusable AI tool wrappers and want an agent to automatically pick the right ones.
• Teams building AI‑assistant products that need to keep token consumption low while still delivering complex multi‑step outputs.
• Anyone looking to reduce the cognitive load of manually mapping tasks to tools.

Where to learn more

• Detailed benchmark data and reproducibility instructions are in the linked vibeskills‑benchmark repo.
• The repository contains case studies (docs/cases/ml-experiment/) and visual overviews of the harness workflow.

---

VibeSkills is essentially a “steward” for your local Skill library, turning a vague user request into a concrete, verifiable, and token‑efficient execution plan.