The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local bash scripts (scripts/list-skills.sh, scripts/validate-skill.sh) and system utilities including git and the claude CLI to perform inventory tasks and run E2E test scenarios.
[DYNAMIC_EXECUTION]: The skill generates new instruction files (SKILL.md) and shell scripts, which are subsequently executed during the E2E testing phase (Phase 5) using claude -p within a temporary directory.
[INDIRECT_PROMPT_INJECTION]: The skill presents an indirect prompt injection surface by ingesting external data from WebSearch and WebFetch during the discovery phase (Phase 0) to inform the creation or improvement of skills. This data is used by a sub-agent to write instruction files that govern agent behavior.
Ingestion points: WebSearch and WebFetch results (Phase 0); conversation history (Phase 1).
Boundary markers: The skill-creator agent is instructed to own only the specific skill directory and is prohibited from editing other skills, agents, or project sources.
Capability inventory: The skill and its sub-agents have access to Read, Write, Edit, Bash, Agent, WebSearch, and WebFetch tools.
Sanitization: No specific sanitization or filtering mechanisms are documented for the external search data before it is processed into new skill instructions.

skills