The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/render-reader-formats.ts executes external commands to render and open documents. On Windows systems, the openFile function uses spawnSync to call cmd /c start with the file path. Because this path is derived from the input filename, a filename containing shell metacharacters (e.g., &, |) could potentially trigger arbitrary command execution, although the script does verify the file exists before proceeding.
[DYNAMIC_EXECUTION]: The skill uses a TypeScript script to dynamically locate and spawn a headless Chrome or Chromium browser at runtime. This process is used to print the generated HTML to a PDF format. The script applies several sandboxing flags (e.g., --headless=new, --disable-gpu, --no-sandbox equivalents) to mitigate risks associated with headless browser execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from a research evidence pack. While this presents an indirect prompt injection surface, the skill instructions include specific 'quality gates' and verification markers ([VERIFY]) to ensure the agent only uses evidenced claims and translates verbatim quotes without inventing facts.

case-business-case