The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill translates arbitrary user requests into executable instruction files (SKILL.md), creating a surface where a malicious user could attempt to generate a skill with harmful procedures.
Ingestion points: User-provided natural language descriptions are used to populate the name, description, and steps of the generated skill file.
Boundary markers: The skill does not define explicit delimiters for user-supplied data during the prompt interpolation process.
Capability inventory: The skill possesses Write capabilities to create filesystem artifacts and Bash to execute verification scripts.
Sanitization: Risk is mitigated by a mandatory validation step that runs validate-skill.sh, which explicitly blocks the inclusion of ~/ paths and verifies YAML frontmatter. It also adheres to a PATTERNS.md checklist at runtime.
[COMMAND_EXECUTION]: The skill executes local shell commands and bundled scripts to discover the workspace structure and verify the integrity of the generated skill.
Executes git rev-parse --show-toplevel to resolve repository and plugin boundaries.
Executes the bundled scripts/validate-skill.sh to perform deterministic checks on the new skill directory.
Executes scripts/lint.sh for repository-level verification when operating within the public catalog.

skill-maker