The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes multiple local scripts through a custom toolchain.
Evidence: It invokes render_page.py, google_auth.py, backlinks_auth.py, drift_history.py, and google_report.py using the claude-seo run command to perform site analysis and report generation.
[PROMPT_INJECTION]: Indirect prompt injection surface is present due to the processing of untrusted external web content.
Ingestion points: The agent crawls up to 500 pages of a user-provided URL and processes the resulting HTML and text as part of the analysis process (SKILL.md).
Boundary markers: No delimiters or instructions to ignore embedded commands are present in the instructions for processing crawled content.
Capability inventory: The agent has capabilities to execute local commands and write files to the filesystem, which could be abused if malicious instructions are ingested.
Sanitization: No evidence of sanitization or filtering for the ingested web content is specified.
[EXTERNAL_DOWNLOADS]: The skill performs extensive network requests to crawl and analyze external web domains.
[CREDENTIALS_UNSAFE]: The skill programmatically checks for the existence of sensitive API credentials.
Evidence: Usage of google_auth.py --check and backlinks_auth.py --check to detect the presence of local Google, Moz, and Bing credentials in order to toggle subagent features.

seo-audit