The Agent Skills Directory

[COMMAND_EXECUTION]: The file examples.md provides an implementation for an execute_code tool that uses the subprocess module to run arbitrary Python and Bash commands on the local system.
[EXTERNAL_DOWNLOADS]: The scripts/check_drift.py utility fetches content from external documentation sources (including Anthropic's official domains and community repositories) using the requests library to monitor for updates.
[DATA_EXFILTRATION]: Implementation templates for filesystem tools in examples.md (safe_file_operation) permit the agent to read and write files based on user-provided or agent-generated paths.
[PROMPT_INJECTION]: The skill architecture is designed for autonomous agents that ingest untrusted data (user tasks) while possessing extensive system capabilities, creating a surface for Indirect Prompt Injection.
Ingestion points: Untrusted user tasks entering the agent context via the Agent.run() method in SKILL.md and examples.md.
Boundary markers: The skill suggests role-based system prompts for instructions but does not enforce hard delimiters for user input by default.
Capability inventory: High-privilege tool capabilities including shell access (bash), file operations, and MCP server integration are documented across reference.md and examples.md.
Sanitization: While examples.md provides a SecurityValidationHook for filtering dangerous command patterns, it is presented as an optional example rather than a required architectural constraint.

claude-agent-sdk