agentsope/SkillAlchemy

SkillAlchemy – Turning Descriptions into Installable Agent Skills

What it is – SkillAlchemy is an open‑source system that automatically builds reusable agent skills from vague task briefs and publicly‑available sources (docs, code repos, papers, issue trackers, etc.). It treats skill creation as a source‑grounded procedure‑admission problem: it discovers missing requirements, extracts executable procedures, judges how broadly each procedure is justified, and packages the result so an LLM‑based agent can load and run the skill directly.

Why it matters – Building reliable skills for autonomous agents is hard because real‑world tasks are often under‑specified and there may be no curated recipe. SkillAlchemy fills that gap by:

• surfacing hidden constraints,
• pulling together evidence‑backed steps from heterogeneous web content,
• separating universally reusable instructions from narrow examples, and
• delivering the outcome as a ready‑to‑install skill package.

Key capabilities

• Implicit requirement discovery – infers constraints and operational dimensions omitted from the original brief.
• People & method distillation – can create Persona skills (capturing decision‑making style, values, communication tone) and Method skills (turning books, tutorials, repositories, or interview transcripts into step‑by‑step procedures).
• Evidence‑based admission – evaluates the scope of each retrieved procedure, keeping only those supported by sufficient public evidence.
• Skill fusion – merges existing workflows or domain knowledge into a new composite capability.
• Installable skill packages – outputs a self‑contained directory (output/) that agents such as Claude Code or Codex can import and execute.

How to get started

# Quick install via the provided CLI (requires Node.js)
npx skills add agentsope/SkillAlchemy

Then ask the agent to create a skill, e.g.:

Use SkillAlchemy to create a Skill for reviewing RAG systems.
Use public documentation and research papers as sources.

The generated package appears under output/. Individual sub‑skills (e.g., Lens, LEAP) can also be installed separately.

Performance – Evaluated on 87 tasks from SkillsBench v1.1 across four agent‑model pairings. SkillAlchemy achieved an average 55.8 % task‑pass rate, beating strong baselines by 8–20 percentage points and matching or slightly surpassing human‑curated skills.

Paper & citation – The full methodology and experiments are described in the arXiv paper SkillAlchemy: Open‑World Agent Skill Creation (2026). See the README for the BibTeX entry.

License – MIT.

---

All details are taken directly from the repository’s README; no additional assumptions have been made.