The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches data from the internet via WebSearch and WebFetch tools and synthesizes it into research reports. This creates a surface for indirect prompt injection where malicious instructions embedded in external web pages could influence the agent's behavior during synthesis. Evidence Chain: 1. Ingestion points: WebSearch and WebFetch tool outputs used in the RETRIEVE and TRIANGULATE phases (SKILL.md). 2. Boundary markers: The instructions advise the agent to distinguish facts from synthesis but do not provide specific delimiters or escape sequences for untrusted content. 3. Capability inventory: The skill utilizes Write, Edit, and Task tools, and executes local Python scripts via the shell. 4. Sanitization: No explicit sanitization or filtering of external content is mentioned before processing.
[COMMAND_EXECUTION]: The skill directs the agent to execute internal Python utility scripts for citation verification, report validation, and HTML conversion. Evidence: SKILL.md contains instructions to run python scripts/verify_citations.py, python scripts/validate_report.py, and python scripts/md_to_html.py.
[EXTERNAL_DOWNLOADS]: The skill performs web searches and content retrieval as its primary function. Evidence: SKILL.md and methodology.md specify the use of WebSearch, WebFetch, and mcp__Exa__exa_search tools to gather information from various domains.
[DYNAMIC_EXECUTION]: The skill uses the Task tool to spawn sub-agents for parallel information retrieval and report auto-continuation. Evidence: SKILL.md describes a Recursive agent spawning system and launching 3-5 parallel agents using the Task tool.

deep-research