The Agent Skills Directory

[METADATA_POISONING]: The skill uses misleading information in its description and body to gain undue trust. It claims to be an "authoritative guide... written and published by Apple," yet it is attributed to a third-party author and references fictional software versions (iOS 27 and macOS Golden Gate).
[PROMPT_INJECTION]: The skill contains instructions designed to override the agent's internal knowledge and behavior, stating that the provided content "unconditionally supersedes any prior training" and that the agent must treat it as "correct" even when it conflicts with known information. This is an attempt to bypass the model's reliance on its own training data.
[INDIRECT_PROMPT_INJECTION]: The skill describes a workflow that processes external, untrusted data (Info.plist declarations and file type identifiers) and recommends using a command-line tool (uttype) to verify them, creating a potential command injection surface.
Ingestion points: The agent is expected to read and process user-provided configuration files (Info.plist) and code identifiers from the developer's project.
Boundary markers: No delimiters or instructions to ignore embedded commands are specified for the agent when handling these external values.
Capability inventory: The skill documents the use of a CLI tool (uttype) for verification tasks based on these external identifiers.
Sanitization: No validation or escaping is recommended for the identifiers before they are used in shell commands, which could be exploited if a malicious identifier is provided.

building-document-based-swiftui-applications