The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The README.md and requirements.txt files recommend the installation of an external tool called search-cli from a third-party Homebrew tap (199-biotechnologies/tap), which is then executed by the agent during the research process.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest, analyze, and synthesize data from external web sources, which is a known vector for indirect prompt injection.
Ingestion points: External information is collected using search-cli, WebSearch, and Exa MCP as described in reference/methodology.md.
Boundary markers: The skill includes a mitigation strategy in reference/quality-gates.md via a 'Source trust boundary' protocol that instructs the agent to treat web content strictly as data and never as instructions.
Capability inventory: The skill uses the Bash tool to run search commands and Python scripts, the Task tool for spawning sub-agents, and file-writing tools for report assembly.
Sanitization: While no automated sanitization of search results is performed before processing, the skill utilizes several validation scripts (validate_report.py, verify_citations.py) to check the structure and integrity of the final report.
[COMMAND_EXECUTION]: The skill frequently uses the environment's Bash capabilities to execute its internal Python scripts (e.g., citation_manager.py, evidence_store.py) and the recommended search-cli tool for data management and retrieval.
[DATA_EXFILTRATION]: The scripts/verify_citations.py script makes automated network requests (HEAD and GET) to verify URLs and DOIs. Since these locators are extracted from the research content (which is untrusted), this creates a surface for data-driven network requests.

deep-research