The Agent Skills Directory

[SAFE]: The skill provides documentation and code examples for official Anthropic SDKs (anthropic for Python and @anthropic-ai/sdk for TypeScript).
[CREDENTIALS_UNSAFE]: The skill explicitly advises against hardcoding sensitive information, demonstrating the correct use of environment variables (e.g., client = anthropic.Anthropic()) and providing placeholders like your-api-key-here for setup instructions.
[COMMAND_EXECUTION]: The skill contains standard installation commands (pip install, npm install) for official packages which are benign and necessary for the stated purpose.
[DATA_EXFILTRATION]: No unauthorized network operations or exfiltration patterns were detected. All network operations described are standard interactions with official Anthropic API endpoints through the provided SDKs.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The patterns shown, particularly the 'Claude Agent SDK' loop, ingest data from user messages and tool results into the model context.
Boundary markers: The code uses standard SDK role markers (user, assistant, tool_result) which provide basic structural boundaries.
Capability inventory: The skill demonstrates model invocation capabilities via API calls.
Sanitization: As these are educational code snippets, explicit input sanitization is not implemented in the examples, but this is typical for SDK documentation.
This represents the standard operational surface for LLM applications rather than a malicious injection attempt within the skill itself.

claude-api