Definition

Agent skills are markdown (.md) documents packaging procedural knowledge for LLM agents — domain heuristics, tool-use policies, output constraints, and known failure modes — inserted into agent context before task execution.

Key Points

  • External interface for adapting frozen models to enterprise workflows
  • Traditionally hand-crafted or one-shot generated without systematic optimization
  • skillopt treats skills as trainable parameters with validation gates
  • Compact artifacts (300–2,000 tokens) portable across harnesses and models

Vercel Skills Ecosystem (2026)

Optimization Challenges

Without systematic training:

  • No step-size control → skill drift
  • No validation → plausible edits regress performance
  • No negative memory → repeated failed edits

Sources