The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell commands to facilitate its workflow.
Uses curl to download the axe-core accessibility library to a local temporary directory (/tmp/screenwright/axe.min.js) in references/playwright-loop.md.
References the execution of npx impeccable detect for performing design quality audits in references/verification.md.
[COMMAND_EXECUTION]: Utilizes high-privilege MCP tools to perform browser-based verification.
Relies on browser_run_code_unsafe to execute complex JavaScript logic within the browser for the verification loop (Phase 4), which includes rendering and inspecting UI surfaces.
[EXTERNAL_DOWNLOADS]: Fetches code and configuration from external repositories.
Downloads the axe-core library from the well-known CDN jsdelivr.net.
Retrieves design templates from a third-party GitHub repository at github.com/VoltAgent/awesome-design-md in references/brand-ladder.md.
[REMOTE_CODE_EXECUTION]: Injects and executes downloaded or generated content within a browser context.
Uses page.addScriptTag to load the downloaded axe-core script into the active page to perform audits.
Utilizes page.setContent to render dynamically generated HTML that may incorporate untrusted user input.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection through its data processing flow.
Ingestion points: Ingests user-provided briefs, mockup images, and existing UI code as described in references/intake.md.
Boundary markers: Lacks explicit boundary markers or instructions to isolate untrusted data from the agent's internal logic during the generation phase.
Capability inventory: Employs powerful capabilities including browser_run_code_unsafe, page.setContent, and file system writes (references/playwright-loop.md, references/handoff.md).
Sanitization: Does not specify sanitization or validation routines for ingested data before it is interpolated into the generation and verification loops.

screenwright