The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content from LinkedIn to generate comments and monitor engagement, creating an attack surface for indirect prompt injection.
Ingestion points: The skill reads external content through the mcp__claude-in-chrome__read_page and get_page_text tools, as well as via curl for public articles.
Boundary markers: There are no explicit instructions or delimiters provided to help the agent distinguish between its system instructions and the untrusted data fetched from LinkedIn posts.
Capability inventory: The agent has the capability to execute JavaScript in a browser environment, generate images via the Gemini API, and write tracking data to a local tracking.md file.
Sanitization: No sanitization, filtering, or validation mechanisms are mentioned for the content retrieved from external sources before it is processed.
[DYNAMIC_EXECUTION]: The skill utilizes a browser automation tool (mcp__claude-in-chrome__javascript_tool) to execute arbitrary JavaScript snippets. While the provided snippets are static templates for data extraction, this mechanism allows for dynamic code execution within the browser context.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of several third-party Python libraries, including google-genai, pillow, and python-dotenv. These are well-known or originate from trusted organizations (Google), but their installation constitutes a dependency on external code.
[CREDENTIALS_UNSAFE]: The skill instructions specify the use of .env.local to store and load sensitive API keys, such as GEMINI_API_KEY. While this is a common development practice, it involves the agent interacting with sensitive configuration files.

linkedin-engagement