The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The configuration in references/mcporter.json specifies the use of npx -y to run @kunihiros/google-patents-mcp and exa-mcp-server. This causes the agent to download these packages from the public npm registry whenever the corresponding tools are invoked.
[REMOTE_CODE_EXECUTION]: By invoking npx -y on remote packages without specifying version pins, the skill executes code that is fetched at runtime from external sources. This lack of version control allows for the execution of potentially compromised or modified code if the upstream packages are altered.
[COMMAND_EXECUTION]: The skill requires the agent to run an initialization shell script (scripts/init.sh) and utilizes shell commands for workspace management (e.g., mkdir, python3 for UUID generation) and tool execution via the mcporter CLI.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: The skill retrieves external data through google-patents-mcp.search_patents and exa.web_search_exa, which is then saved to files such as search-report.md and similar-patents-reference.md for processing by subsequent sub-agents.
Boundary markers: The instructions do not define delimiters or protective prompts to prevent the agent from obeying malicious instructions that might be embedded in the retrieved patent text or web content.
Capability inventory: The agent environment permits shell command execution via mcporter and bash, as well as file system modifications within the work directory.
Sanitization: There is no evidence of sanitization, filtering, or validation performed on the search results before they are passed into the prompts of the writing agents.

patent-draft-agent