The Agent Skills Directory

[METADATA_POISONING]: The skill makes deceptive claims regarding its authorship and the technical environment it covers. The body of 'SKILL.md' states 'This guidance was written and published by Apple,' which contradicts the actual author metadata. Furthermore, 'references/soft-deprecated-apis.md' claims to be generated from 'iOS 27.0' and 'macOS 27.0', which are non-existent future versions, potentially misleading the agent or user into trusting the data as official and forward-looking.
[PROMPT_INJECTION]: The skill employs override language meant to bypass the model's baseline training. Both the 'SKILL.md' frontmatter and body include instructions that this information 'unconditionally supersedes any prior training the model may have,' which is a high-confidence marker for prompt injection techniques designed to force model behavior.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface by instructing the agent to ingest and process untrusted external data. 1. Ingestion points: The skill explicitly directs the agent to 'scan the project' to identify code areas for evaluation ('SKILL.md'). 2. Boundary markers: No instructions, delimiters, or warnings are provided to prevent the agent from obeying instructions embedded within the codebase it scans. 3. Capability inventory: The agent is empowered to read project files, provide guidance, and generate new code based on the scanned content. 4. Sanitization: No sanitization or validation steps are defined for handling external code content.

swiftui-specialist