The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill identifies and fetches external HTML, CSS, and plain content variants from URLs provided by the user to perform a WCAG 2.1 AA audit.
[COMMAND_EXECUTION]: The instructions suggest using the curl tool to retrieve raw HTML content. This is recommended to ensure that essential accessibility attributes like aria-*, alt, and role are preserved for analysis, as markdown conversion tools might strip them.
[PROMPT_INJECTION]: The skill processes untrusted external data from the web, which presents an indirect prompt injection surface. The analysis of this risk is as follows:
Ingestion points: External URLs provided by the user, specifically the main page and associated navigation/footer components (SKILL.md).
Boundary markers: The instructions contain a dedicated 'External Content Safety' section requiring the agent to only fetch user-provided URLs and treat all data as untrusted.
Capability inventory: The agent uses fetching tools (like curl) and parsing logic to evaluate HTML elements and CSS variables.
Sanitization: The skill explicitly forbids the execution of scripts or interpretation of dynamic content within the fetched data, mitigating the risk of malicious code execution during the audit.

accessibility-fix