The Agent Skills Directory

[PROMPT_INJECTION]: The skill facilitates a meta-development workflow that ingests and processes untrusted input from user-provided requests and existing skill files, creating an indirect prompt injection surface.
Ingestion points: Untrusted data enters the agent context through the request input and by reading existing SKILL.md files as defined in the optimization flow in references/workflow-playbook.mdx.
Boundary markers: The workflow includes a mandatory dry-run by default and requires explicit user confirmation before applying any changes, providing a human-in-the-loop safety check.
Capability inventory: The skill has the capability to write and modify files on the local filesystem using the skill-creator-create and skill-creator-optimize tools.
Sanitization: The execution rules in SKILL.md specify that the skill must validate the scope and block any invalid or unsafe write targets to prevent unauthorized file access.

skill-creator