The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted content from external websites, creating a surface for indirect prompt injection.
Ingestion points: Website content is fetched via WebFetch and crawled using a site-mapper agent.
Boundary markers: Sub-agents are explicitly instructed to return results only in JSON format to prevent prose-based injection from leaking into the final report.
Capability inventory: The environment includes tools for file system access (Read, Write), network access (WebFetch), and shell execution (Bash, PowerShell).
Sanitization: The skill performs basic sanitization by stripping markdown fences from agent responses before JSON parsing.
[EXTERNAL_DOWNLOADS]: The skill uses a package manager to fetch auditing utilities at runtime.
Fetches axe-cli, pa11y, Lighthouse, and @playwright/mcp from the official NPM registry using npx.
[COMMAND_EXECUTION]: Shell commands are used for environment detection and workspace management.
Executes uname -s to identify the operating system.
Uses mkdir -p (Unix) or New-Item (PowerShell) to create a dedicated directory for audit artifacts in system temporary locations.
[DYNAMIC_EXECUTION]: The skill dynamically generates testing scripts based on site observations.
The bug-script-runner agent writes a Playwright test specification (bugs.spec.ts) tailored to the specific selectors found on the target website and subsequently executes it.

audit-site