The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a set of vendor-provided Python scripts located in ${CLAUDE_PLUGIN_ROOT}/scripts/claude-seo to perform its audits. These include lighthouse_agentic.py, agentic_check.py, agent_ux_check.py, and render_page.py. These scripts are executed via shell commands to process site data and generate JSON reports.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from external URLs, including site markup, robots.txt, llms.txt, and ai-catalog.json.
Ingestion points: Fetched site content and discovery files via agentic_check.py and lighthouse_agentic.py.
Boundary markers: The instructions in SKILL.md under the 'Security' section explicitly mandate that fetched content must be treated as untrusted data and never as instructions. It specifically notes that discovery files addressing the agent should be reported as findings, not followed as commands.
Capability inventory: Subprocess execution of the claude-seo script suite for auditing; drafting of configuration files to stdout (no automated file writes to the host).
Sanitization: All network requests are routed through a url_safety wrapper to prevent SSRF and DNS-rebinding attacks.
[EXTERNAL_DOWNLOADS]: The skill fetches content from user-provided target URLs for auditing purposes. It also references and may fetch configuration/IP data from well-known services such as Google, OpenAI, Anthropic, and Cloudflare to verify bot identities. These fetches are performed using a security-hardened utility.
[DATA_EXFILTRATION]: In the WAF behavior check mode (--ua-matrix), the skill sends network requests carrying AI agent User-Agent tokens to the target URL to observe server responses. This is a functional requirement of the audit and is constrained by the url_safety guardrails to prevent misuse against internal or restricted hosts.

seo-agentic