The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references and provides installation instructions for official Anthropic SDKs (e.g., claude-agent-sdk, @anthropic-ai/sdk) and MCP servers. These downloads target official package registries and trusted GitHub repositories belonging to the Anthropic organization.
[COMMAND_EXECUTION]: The documentation describes the use of powerful built-in tools such as 'Bash' and 'code_execution'. It includes appropriate security guidance, such as recommending the use of os.path.basename() to prevent path traversal attacks when handling generated files.
[SAFE]: Instructions for handling sensitive data correctly point users toward environment variables (e.g., ANTHROPIC_API_KEY, DATABASE_URL) rather than hardcoding credentials.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the development of agents that ingest untrusted external data (via web search or file reading) and have the capability to execute commands or write to the filesystem. The documentation mitigates this risk by highlighting safety features, permission modes (e.g., default prompting for dangerous operations), and the necessity of input validation.
Ingestion points: Project files processed for language detection in SKILL.md; web content retrieved via WebFetch in shared/live-sources.md; user-provided files analyzed in code execution patterns.
Boundary markers: The documentation refers to Anthropic's native message structures and beta features designed to manage context boundaries.
Capability inventory: The skill documents tools for file reading (Read, Glob), file modification (Write, Edit), and shell execution (Bash) across various script examples.
Sanitization: Explicit guidance is provided in shared/tool-use-concepts.md and python/claude-api/tool-use.md to use path sanitization functions like os.path.basename().

claude-api