Agent Skill

Agent Skill is a reusable, structured procedural artifact that guides an AI agent through a class of tasks without retraining the underlying model.

Why It Exists

Agents repeatedly encounter the same procedures, tool semantics, and failure modes. A skill stores that reusable knowledge so a later run can start with tested guidance instead of rediscovering it through trial and error.

How It Works

A skill can be written by a person or extracted from successful and failed interaction trajectories. It normally contains a short description, procedural instructions, applicability conditions, pitfalls, and optional references or scripts. At inference time it is either injected into the agent prompt or retrieved from a library when the task matches its scope.

The important validation target is behavior: a skill is useful when it improves held-out task performance for the intended consumer. Fluent wording, a tidy format, or broad-sounding advice is not sufficient; a skill can encode an unsafe shortcut or a procedure that a particular model cannot execute.

Tradeoffs

Skills reduce repeated reasoning and enable fast adaptation, but they can transfer errors and create interference. A skill extracted from a weak experience pool may cause negative transfer, while a skill that helps one model may be neutral or harmful for another. Large libraries add retrieval, composition, and versioning problems that a single prompt-injected skill does not expose.

Common Confusions

• Agent skill vs. tool: A tool performs an operation; a skill explains when and how to combine operations for a class of tasks.
• Agent skill vs. memory: Memory records past information or events; a skill distills reusable procedure and decision rules.
• Agent skill vs. fine-tuning: A skill changes the runtime context without changing model parameters; fine-tuning changes the parameters or learned weights.

Where It Appears

• From Raw Experience to Skill Consumption — Studies the complete experience-generation, extraction, and consumption lifecycle and validates a failure-aware extraction rubric against downstream utility.
• SkillOpt: Executive Strategy for Self-Evolving Agent Skills — Treats one skill document as the trainable state of a frozen agent and validates bounded text edits on held-out tasks.
• Socratic-SWE — Distills coding-agent traces into a registry that guides targeted task generation in a self-evolution loop.
• AutoSkill: Experience-Driven Lifelong Learning — Extracts user-side behavioral rules into versioned SKILL.md artifacts and retrieves them for later requests without updating model weights.