The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches live HTML content from user-provided URLs using tools like WebFetch and /browse. This external content (metadata, headings, body text, and schema) is directly processed by the agent to generate scores and recommendations. This creates a surface for indirect prompt injection, where a malicious website could embed instructions designed to influence the audit results or the agent's behavior.
Ingestion points: WebFetch of $PAGE_URL and competitive URLs in Phases 1a and 5.
Boundary markers: Absent; the agent is instructed to extract and evaluate content directly from the raw HTML without specific isolation instructions.
Capability inventory: Execution of local Python scripts (Phase 1c-e), tool invocation (WebFetch, WebSearch), and file system reads (Phase 1f).
Sanitization: The instructions do not specify any sanitization or filtering of the fetched content before analysis.
[COMMAND_EXECUTION]: The skill uses bash blocks to execute local Python scripts (e.g., analyze_gsc.py, list_gsc_sites.py, url_inspection.py) and standard utilities like python3 -c for URL parsing. These operations handle data derived from user-provided URLs.
[DATA_EXPOSURE]: The skill attempts to read business context data from the local file system at $HOME/.toprank/business-context/$DOMAIN.json. While this is part of the intended functionality, it involves reading files from the user's home directory based on processed domain strings.

seo-page