Agent Skills Turn Reusable Professional Know-How Into Structured Directories, Not Just Saved Prompts
Ad-hoc prompts produce inconsistent results across team members and sessions. Skills give an agent a repeatable professional method that can be version-controlled, tested, and shared, turning scattered chat-history know-how into a maintainable capability.
An Agent Skill is a directory containing a SKILL.md file with YAML frontmatter (name and description) and a Markdown body that defines the workflow, output format, and quality boundaries for a specific task. The description determines when the skill triggers, while the body encodes the professional process — for example, cleaning filler words from a transcript, grouping points by topic, and extracting action items with responsible persons and deadlines. Complex skills add subdirectories for executable scripts, reference documents, and deliverable assets, all loaded on demand to keep token usage low.
Progressive loading is the mechanism that makes dozens of skills practical: the agent first sees only the name and description, loads the full SKILL.md once a task matches, and pulls in scripts or references only when a step requires them. This avoids flooding the context window with irrelevant rules and keeps the agent responsive.
The distinction from related concepts matters. Prompts define a single job; Memory stores long-term facts; MCP and Tools provide connections and operations. A Skill supplies the stable, repeatable method — what to delete, what to keep, how to handle missing data, and what structure to deliver. Tasks worth packaging are those that recur, have stable steps and output formats, carry quality standards that are easy to overlook, and need to be applied consistently across a team.
Progressive loading is the architectural decision that makes large skill libraries feasible — without it, every installed skill would burn tokens and pollute the context window on every request.
The description field carries an outsized burden: it is the sole mechanism for intent matching, yet writing it too narrowly causes missed triggers and too broadly causes skill collisions, with no runtime feedback loop to tune it automatically.
Calling a Skill 'a saved prompt' undersells it. A prompt tells the agent what to produce this time; a Skill encodes the professional judgment rules — edge-case handling, quality gates, missing-data policies — that separate a novice output from an expert one.
Skills, Memory, MCP, and Tools form a capability stack where each layer solves a distinct problem, and conflating them leads to brittle agent designs that either overload the prompt or under-specify the process.