The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external, potentially untrusted artifacts (HTML, screenshots, PR diffs, specifications) which could contain hidden instructions intended to manipulate the behavior of the review personas.
Ingestion points: The skill accepts multiple file paths in Step 1, including screenshot text dumps, HTML captures, markdown specs, and PR diffs.
Boundary markers: Step 3 implements an isolation invariant where subagents are prompted only with artifact paths and a defined stance, specifically avoiding operator summaries to prevent design-aware contamination.
Capability inventory: The skill dispatches parallel subagents and, in 'Explore mode', generates navigation commands for execution by external runners such as Playwright or adb.
Sanitization: No explicit sanitization or filtering of the artifact content is mentioned before it is processed by the subagents.
[SAFE]: The skill documents a reference to an external repository (github.com/wshobson/agents) for role vocabulary, but this is a static attribution and does not involve runtime downloads or execution of remote code. Furthermore, the skill explicitly mandates local processing for sensitive artifacts to prevent data leakage.

persona-cast