The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes patterns for creating specialized agents (e.g., code-reviewer, test-generator) that ingest and process code, which represents a surface for indirect prompt injection attacks.
Ingestion points: Agent definitions in examples/standard-plugin.md and examples/advanced-plugin.md process user-provided code and files.
Boundary markers: While the examples focus on structure, they do not explicitly demonstrate the use of XML delimiters or similar markers in the prompt implementation guidelines.
Capability inventory: The skill documentation covers the execution of shell scripts and commands via the hooks and commands features.
Sanitization: The documentation provides a best-practice example of a safety hook in advanced-plugin.md that uses a prompt-based evaluation to check the safety of bash commands before they are executed.
[COMMAND_EXECUTION]: The skill provides detailed guidance and scripts for executing local tools and commands within the plugin environment.
Evidence: The standard-plugin.md example includes a validate-commit.sh script that runs eslint and pylint on local files.
Best Practices: The skill consistently enforces the use of the ${CLAUDE_PLUGIN_ROOT} environment variable to ensure all script and file references are portable and confined to the plugin's own directory structure.

plugin-structure