The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes GEO-AUDIT-REPORT.md, which is external data typically generated from a website audit. This creates a surface for indirect prompt injection where malicious content in the audit report could influence the agent's behavior during PDF generation.
Ingestion points: GEO-AUDIT-REPORT.md is read to extract metadata and content for the report.
Boundary markers: The skill instructs the agent to treat report fields as untrusted values and never as executable code, but no explicit delimiters are defined for the file content itself.
Capability inventory: The skill uses Bash to execute Google Chrome and a Python script to execute pandoc via subprocess.run.
Sanitization: The embedded Python script performs basic metadata validation, checking keys against an allowed list, ensuring values are strings, and limiting value length to 500 characters.
[COMMAND_EXECUTION]: The skill executes external binaries (pandoc and Google Chrome) using shell commands and Python's subprocess module. While the Python execution uses shell=False, the Chrome command is a direct bash execution.
[DYNAMIC_EXECUTION]: The workflow involves generating a temporary HTML file (GEO-REPORT.html) from untrusted markdown data and then rendering that HTML using a headless browser. This pattern can be exploited if malicious HTML or JavaScript is successfully injected into the intermediate file, although the skill includes instructions to inspect the HTML and disable network access.

geo-report-pdf