The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted HTML and content from user-provided external URLs to generate marketing strategies, articles, and social media posts. This creates a surface for indirect prompt injection where malicious instructions embedded in a website could influence the agent's behavior.
Ingestion points: scripts/site_audit.py crawls the target website, and WebFetch is used in SKILL.md (Step 2b) to pull raw HTML from key pages.
Boundary markers: Absent. The skill does not explicitly instruct the agent to ignore or delimit instructions found within the crawled content.
Capability inventory: subprocess.run (executing Lighthouse), WebSearch, WebFetch, and AskUserQuestion.
Sanitization: scripts/site_audit.py uses a standard HTMLParser to extract specific metadata (titles, tags, headings), which provides basic structural isolation but does not filter the semantic content processed by the agents.
[COMMAND_EXECUTION]: The scripts/site_audit.py script executes the lighthouse or npx lighthouse CLI tools to perform browser-based performance audits. The command is executed using a list of arguments via subprocess.run, which safely handles the user-provided URL without involving a shell.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to arbitrary external domains provided by the user to crawl content. It also makes requests to the Google PageSpeed Insights API (googleapis.com) to retrieve performance metrics.
[DATA_EXFILTRATION]: While the skill does not exfiltrate data to unknown third parties, it handles the GOOGLE_PSI_API_KEY provided by the user or found in the environment to interact with Google's APIs.

cmo-agent