The Agent Skills Directory

[METADATA_POISONING]: The skill provides misleading metadata regarding platform support and official documentation sources. It claims compatibility with iOS 26, macOS 26, and visionOS 26, versions that do not currently exist, and cites a WWDC2025 session which is also in the future relative to current Apple releases.
[EXTERNAL_DOWNLOADS]: The skill references documentation and video resources from the domain sosumi.ai. This is an untrusted third-party domain (known historically as a parody site) rather than the official Apple developer portal (developer.apple.com), posing a risk of misinformation or redirection to untrusted content.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by defining workflows for ingesting untrusted external data into the agent's context without sanitization or boundary markers.
Ingestion points: Data is loaded from external files using PaperMarkup(dataRepresentation: data) in both SKILL.md and references/paperkit-patterns.md.
Boundary markers: Absent. There are no instructions or delimiters to distinguish between system commands and data contained within the markup files.
Capability inventory: The skill uses capabilities for file system writes (data.write), search indexing via indexableContent, and image rendering with markup.draw.
Sanitization: Absent. Content is loaded directly into the data model and processed without validation or filtering of embedded text or metadata.

paperkit