The Agent Skills Directory

External Package Installation & Command Execution: The skill instructions direct the agent to install external Node.js packages (simple-icons, sharp) via the npm package manager and execute shell commands for document processing (soffice.py, pdftoppm).
Evidence: Step 6 and Step 8 in SKILL.md contain commands for installing dependencies and performing QA using system-level tools.
Dynamic Code Execution: The skill provides JavaScript code templates for brand logo retrieval and generation which are intended for runtime execution by the agent.
Evidence: Step 6 in SKILL.md includes logic for identifying brand icons and generating fallbacks using the sharp library.
Indirect Prompt Injection Surface: The skill ingests data from external financial tools (S&P Global Capital IQ) and incorporates summaries directly into the final report slide. While the data is structured, the lack of explicit sanitization or boundary markers for company descriptions presents a potential surface for indirect instruction injection.
Evidence Chain:
Ingestion points: get_rounds_of_funding_info_from_transaction_ids and get_company_summary_from_identifiers in SKILL.md (Steps 3 & 4).
Boundary markers: No explicit delimiters or instructions to ignore embedded commands are defined for external data ingestion.
Capability inventory: The skill utilizes file system writes, Node.js script execution, and shell-based document conversion tools.
Sanitization: No specific sanitization or validation logic is specified for the descriptive text retrieved from external summaries.

funding-digest