What are agent skills?

Agent skills are portable packages of procedural knowledge that teach an AI agent how to perform a specific, repeatable task. A skill can include instructions, executable scripts, reference material, templates, and other resources that the agent loads when the task requires them.

Under the open Agent Skills specification, a skill is a directory with a required SKILL.md file. That file contains YAML metadata, including a name and description, followed by Markdown instructions. Optional scripts, references, and assets live beside it.

A skill is different from a tool. A tool gives the agent an action such as search-customer or create-task. A skill explains when and how to use tools, what sequence to follow, which standards to apply, what output to produce, and how to handle edge cases.

Prompts copied between chats are hard to maintain, test, and distribute. Skills turn a repeatable process into a named package that can be versioned, reviewed, shared, and improved without adding every instruction to every agent session.

Skills also reduce context pressure through progressive disclosure. The agent sees compact metadata for available skills, loads the full SKILL.md only when a skill matches the task, and opens scripts or references only when the instructions require them.

For a company, this creates a practical way to capture operating knowledge. A good skill can preserve a reporting method, document standard, analysis checklist, or customer workflow while still leaving tool permissions and final approvals under the agent runtime's controls.

At session start, the agent receives a catalog containing each skill's name and description. The description carries the activation signal, so it must state both what the skill does and when it should be used.

When a request matches, the agent loads the skill's SKILL.md. It follows the instructions, reads referenced files as needed, and may run bundled scripts if the runtime allows code execution. Resources are resolved relative to the skill directory, which keeps the package portable.

The skill does not grant new permissions by itself. Any file access, code execution, network request, or MCP tool call still depends on the host agent's sandbox, tool configuration, credentials, and approval policy.

A revenue team creates an account-brief skill. Its description says to use it before renewal and executive meetings. SKILL.md defines the required sections, source hierarchy, freshness checks, and the rule that every risk must link to evidence.

The skill includes a script that normalizes dates and a reference file with the company's opportunity-stage definitions. When a seller asks for a renewal brief, the agent loads the instructions, queries the approved CRM tools, runs the date script, and opens the stage reference only when it classifies the opportunity.

If the account lacks a current renewal date, the skill does not invent one. It marks the field as missing, links the source records it checked, and asks the owner to confirm the date before the final brief is shared.

Keep one skill focused on one repeatable outcome. Write a precise description, define inputs and output shape, include decision rules and edge cases, and move long reference material out of SKILL.md so the agent loads it only when needed.

Treat bundled scripts and resources as software supply-chain inputs. Review their source, dependencies, licenses, network behavior, file access, and secret handling. Pin or review versions before organization-wide distribution, especially when a skill can execute code.

Evaluate skills with real tasks. Test correct activation, false activation, output quality, tool selection, missing context, unsafe instructions, and behavior after updates. Record an owner, version, compatibility requirements, allowed tools, and a deprecation path.