agentsope/SkillAlchemy

What it solves

Creating reliable skills for AI agents is often difficult because task descriptions are usually underspecified, expert procedures may not be documented, and execution traces are often missing. SkillAlchemy solves this by automating the creation of installable, reusable agent skills from vague briefs and open-world data sources.

How it works

SkillAlchemy treats skill creation as a source-grounded procedure-admission problem. It follows a four-step process:

1. Requirement Discovery: It identifies implicit requirements and constraints missing from the initial skill brief.
2. Procedure Acquisition: It gathers grounded procedures from public sources like documentation, repositories, papers, and issue reports.
3. Scope Determination: It evaluates evidence to decide if a procedure is a reusable instruction, a specific example, or should be excluded.
4. Compilation: It packages the admitted knowledge into an installable Skill package that agents can load and use directly.

Who it’s for

It is designed for developers and researchers building AI agents who need to equip their agents with specialized capabilities (such as reviewing RAG systems) without manually authoring every step of the workflow.

Highlights

• Implicit Requirement Recovery: Recovers behavior-relevant constraints and operational dimensions omitted in briefs.
• Knowledge Distillation: Converts people (personas), methods (books, papers, repositories), and existing workflows into executable skills.
• Evidence-Based Filtering: Separates reusable instructions from context-specific examples to ensure reliability.
• Skill Fusion: Ability to combine multiple existing workflows or domain knowledge sets into a single new capability.
• High Performance: Outperforms several automated skill-creation baselines and performs comparably to human-curated skills on the SkillsBench v1.1 benchmark.