The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to audit arbitrary external websites, creating an inherent surface for indirect prompt injection.
Ingestion points: Data is collected from external URLs via scripts/shoot.mjs (screenshots, console errors, and accessibility reports) and scripts/audit.mjs (extracted page text and metadata).
Boundary markers: The 'Critic brief' in references/quality-bar.md provides instructions to the subagent to restrict its focus ('Use only the files below: don't open the site, search the web, or read anything else about it'), which acts as a logical boundary.
Capability inventory: The skill uses subagent invocation to process evidence and performs file writing to the .jury/ directory to store audit results.
Sanitization: The audit.mjs script performs basic sanitization by stripping HTML tags from the page content before presenting it to the critic, though visual information in screenshots remains unsanitized.
[EXTERNAL_DOWNLOADS]: The scripts/shoot.mjs script attempts to download the axe-core accessibility library from a well-known service (jsDelivr CDN) at https://cdn.jsdelivr.net/npm/axe-core@4/axe.min.js if the package is not found in the local environment. This is a standard dependency for the skill's accessibility auditing functionality.

jury