The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits a vulnerability surface for indirect prompt injection during the creation and planning phases in Steps 1 and 2 of SKILL.md. It ingests untrusted user-provided examples and requirements to scaffold new skill files without specifying sanitization protocols.
Ingestion points: User-provided examples and functionality requirements gathered in SKILL.md (Steps 1 and 2).
Boundary markers: The instructions lack guidance on using delimiters or "ignore embedded instructions" warnings when incorporating user examples into generated skill files.
Capability inventory: The skill utilizes Write, Edit, Bash, and Agent tools throughout its workflow, providing significant execution capabilities to generated skills.
Sanitization: No sanitization, escaping, or validation of external content is mandated before interpolation into the new skill's SKILL.md or resource files.
[PRIVILEGE_ESCALATION]: The references/tools.md file provides technical guidance on using the dangerouslyDisableSandbox parameter for the Bash tool. Documentation of parameters that bypass security sandboxes increases the risk of privilege escalation if implemented in created skills.
[EXTERNAL_DOWNLOADS]: The references/scripts.md file includes instructions for developers to specify installation commands for external dependencies from well-known registries. It notes that the platform environment supports fetching packages from npm, PyPI, and GitHub repositories. These references are documented as standard platform capabilities.
[COMMAND_EXECUTION]: The skill facilitates the execution of arbitrary shell commands. SKILL.md Step 8 directs the agent to run project-specific test suites via the Bash tool. Additionally, references/tools.md provides detailed reference for Bash tool parameters including timeout configurations and background execution.

create-skill