The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The configuration in references/mcporter.json instructs the agent to execute code from a remote source via npx -y @kunihiros/google-patents-mcp. This package originates from a third-party personal scope on the NPM registry and is executed without version pinning, which poses a risk of executing unverified or potentially malicious code. It also fetches the exa-mcp-server from the well-known search service Exa.
[EXTERNAL_DOWNLOADS]: The skill's operational logic relies on external software components fetched at runtime or installation, including mcporter and the markitdown document conversion tool from Microsoft.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied technical materials (DOCX, PDF, Markdown) which may contain malicious instructions designed to manipulate agent behavior.
Ingestion points: Technical disclosure files provided by the user and processed via the input-parser prompt.
Boundary markers: No delimiters or specific instructions were identified to isolate the LLM from commands embedded within technical source text.
Capability inventory: The agent has capabilities to execute shell commands (markitdown, bash, mkdir), write to the local file system, and perform network operations via MCP.
Sanitization: No content validation or sanitization is performed on the technical content before it is interpolated into subsequent prompts.
[COMMAND_EXECUTION]: The skill executes shell commands for document processing, such as markitdown <input-file> -o <output-file>. Using user-provided file names as arguments to shell commands creates a potential injection surface if the execution environment does not strictly sanitize inputs.

patent-writer-agent