The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill requires the installation of the claude-agent-sdk Python package and the claude-code NPM package, both of which are official tools from Anthropic.
[COMMAND_EXECUTION]: The verification script and the underlying SDK transport execute the claude-code CLI using system subprocess calls for agent communication.
[PROMPT_INJECTION]: The skill defines an agent development framework that inherently possesses an indirect prompt injection vulnerability surface.
Ingestion points: Data enters the agent context through the prompt parameter in query() and ClaudeSDKClient methods across the provided code examples and scripts.
Boundary markers: The skill documentation (SKILL.md, implementation-guide.md) encourages the use of system_prompt and project context (CLAUDE.md) to establish agent constraints.
Capability inventory: Agents built using this skill are typically granted access to system tools like Bash, Write, and Read as specified in the metadata.
Sanitization: The skill provides comprehensive defensive examples, including dynamic can_use_tool callbacks and PreToolUse hooks (in hook_integration_example.py and implementation-guide.md) specifically designed to filter or block malicious operations like 'rm -rf'.

agent-sdk-python