The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes external UI, HTML, and accessibility trees, which are untrusted data sources. A malicious page or prototype could attempt to influence the agent's critique through hidden instructions embedded in the design content.
Ingestion points: The skill ingests data from captured UI, HTML source, accessibility trees, and network request logs (method, URL, and body) from provided URLs or local builds as described in references/sources.md.
Boundary markers: The skill does not explicitly define delimiters for UI text but enforces strict adherence to the criteria defined in the local references/review-criteria.md file to mitigate opinion-based findings.
Capability inventory: The skill has the capability to write report files and screenshots to the .design-review/ directory, execute local scripts like node scripts/check-spec.mjs, and intercept network traffic using browser automation tools.
Sanitization: No specific sanitization or escaping of captured UI text is mentioned before the agent processes it for the review report.
[COMMAND_EXECUTION]: The skill instructions include executing local scripts to perform specific audits.
Evidence: SKILL.md and sources.md instruct the agent to run node scripts/check-spec.mjs to verify component specifications against the rendered UI.

design-review