SkillOpt — Train Reusable Natural-Language Skills for LLM Agents

Introduction

SkillOpt is a Microsoft Research project that trains reusable natural-language skills for frozen LLM agents. Rather than fine-tuning weights, it iteratively refines a text-space skill document that an agent loads into its context to improve task performance. The result is a deployable best_skill.md artifact any compatible agent can consume.

What SkillOpt Does

• Optimizes natural-language skill documents through trajectory-driven editing
• Evaluates agent trajectories on validation tasks and proposes targeted revisions
• Produces a standalone best_skill.md file for any frozen LLM agent
• Supports multiple task families for cross-task generalization
• Gates every update on validation performance to prevent regressions

Architecture Overview

SkillOpt runs a three-phase loop: execute, analyze, revise. The agent runs tasks with the current skill, SkillOpt examines trajectories for failure patterns, and a proposer model drafts edits that are accepted only if validation scores improve. Written in Python, it requires only API access to the target LLM.

Self-Hosting & Configuration

• Install via pip from the cloned repository
• Provide an LLM API key via environment variable
• Define task suites as JSON with inputs and evaluation criteria
• No GPU required — operates entirely through API calls

Key Features

• Learns from actual agent trajectories rather than static benchmarks
• Validation-gated updates ensure measurable improvement on every revision
• Produces portable markdown skill files usable across agent frameworks
• Open-source under MIT with example task suites included

Comparison with Similar Tools

• DSPy — compiles prompts; SkillOpt optimizes skill documents via trajectory analysis
• TextGrad — gradient-like text feedback; SkillOpt uses validation-gated discrete edits
• OPRO — searches prompt variations; SkillOpt refines structured skills from trajectories

FAQ

Q: Does it require fine-tuning the LLM? A: No. The LLM stays frozen; only the text skill document is modified.

Q: What is the output format? A: A markdown file (best_skill.md) with structured instructions and heuristics.

Q: How many iterations does optimization take? A: Most tasks converge within 10 to 30 iterations depending on complexity.

Sources

• https://github.com/microsoft/SkillOpt