What are Agent Skills? Skill files, scripts, and on-demand loading
Understand reusable skill packages through a report-checking example, distinguishing prompts, tools, training, progressive loading, dependencies, and permissions.
Understand the concept
Agent Skills package reusable instructions and resources in directories that compatible agents load as needed. The open format uses SKILL.md and can include scripts and references. Skills do not retrain models or grant arbitrary permissions. Use depends on host support, dependencies, and execution authorization.
Load entry instructions and detailed resources as needed
A report-checking skill can state scope and procedure in its entry and keep field rules in references. The agent can inspect names and descriptions before loading details. Progressive disclosure reduces initial context but does not guarantee correct selection.
The Agent Skills specification defines structure and metadata, and Anthropic describes loading levels. Hosts differ; copying a folder does not establish portability. Check paths, dependencies, and any tools scripts require.
Instructions and executable tools play different roles
A skill can guide header and amount checks, while tools or programs actually read the spreadsheet. Tools provide actions; skills describe task methods. Missing execution dependencies should be reported rather than replaced by invented check results.
Bundled scripts are not inherently trusted; verify source and relevant operations. Instructions to send externally or delete do not expand user authorization. Installation and approval for a consequential action are different; application controls still enforce permissions.
Validate actual outputs rather than the skill’s name
Use a sample report with known errors and check whether positions are identified without changing correct fields. This is a validation design, not a skill run. Successful loading does not establish complete dependencies or correct conclusions.
Track versions and environments and recheck tasks affected by changed rules. Skills reduce repeated instructions but do not confer credentials or permanent attachment memory. Distinguish documentation review from execution so evidence remains clear.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.