The Agent Skills Directory

[PROMPT_INJECTION]: The implementation patterns create a surface for indirect prompt injection by processing untrusted user messages while granting agents access to sensitive capabilities.
Ingestion points: The invoke method in src/ai/agent-executor.ts processes a messages array which typically contains external, untrusted content.
Boundary markers: No boundary delimiters or explicit 'ignore instructions' directives are implemented in the provided agent templates to isolate user data from system instructions.
Capability inventory: The framework allows registering high-privilege MCP tools such as playwright for browser automation and @modelcontextprotocol/server-filesystem for workspace access.
Sanitization: There is no evidence of input validation or sanitization of the content passed into the agent's context.
[SAFE]: The skill utilizes official packages from trusted organizations and well-known services.
Fetches the Claude SDK and provider-specific libraries from Anthropic's official repositories.
Recommends well-known and maintained MCP tools from the Playwright and Model Context Protocol organizations.
Follows security best practices by using environment variables for credential management rather than hardcoding secrets.

claude-agent-sdk