The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use bash_tool to execute curl commands for querying the NIH RePORTER API. The instructions provide templates that interpolate search terms directly into a JSON payload within a shell command, which poses an argument injection risk if user-supplied or externally fetched search terms contain shell metacharacters.\n- [DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file uses the !command`` syntax to execute internal utility scripts, such as scripts/fiscal_year_calculator.py, during the skill load phase. While used for benign environment setup (calculating the current federal fiscal year), this demonstrates the use of load-time dynamic context injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, including Consensus paper summaries, NIH RePORTER project abstracts, and grant notice pages, creating a vulnerability surface for indirect prompt injection.\n
Ingestion points: SKILL.md (Phases 2A and 2B) details fetching findings from Consensus, project metadata from RePORTER, and NOSI text from NIH.gov.\n
Boundary markers: There are no explicit instructions or delimiters defined to prevent the agent from following instructions that might be embedded within the retrieved research abstracts or grant notices.\n
Capability inventory: The skill has access to bash_tool for command execution and network operations, Python for data processing, and Node.js for file generation (.docx).\n
Sanitization: No evidence of sanitization, validation, or escaping of external content is provided before it is interpolated into prompts or documents.\n- [EXTERNAL_DOWNLOADS]: The skill fetches research and grant data from established and well-known services, including the official NIH RePORTER API (api.reporter.nih.gov) and consensus.app. These references are documented as part of the skill's core functionality for clinical researchers.

grants