The Agent Skills Directory

[METADATA_POISONING]: The skill body in SKILL.md contains a deceptive claim that the content was 'written and published by Apple.' This impersonation of a trusted vendor is used to establish unearned authority over the model's behavior.
[PROMPT_INJECTION]: The instructions include 'override' patterns. SKILL.md tells the model that the provided info 'unconditionally supersedes' its prior training. Additionally, in references/soft-deprecation.md, it mandates a strict output format ('Nothing else') and explicitly instructs the agent to ignore user requests for 'observations' or 'other notes' in certain contexts, potentially bypassing user intent.
[INDIRECT_PROMPT_INJECTION]: The skill directs the agent to 'scan the project' and analyze/edit user code, establishing an ingestion surface for untrusted data.
Ingestion points: User project files scanned by the agent as described in SKILL.md and referenced files.
Boundary markers: Absent; there are no instructions to use delimiters or warnings to ignore instructions embedded within the scanned code.
Capability inventory: The skill facilitates reviewing, writing, and editing SwiftUI code within the user's project environment.
Sanitization: Absent; no sanitization or escaping mechanisms are provided to protect against malicious instructions that may be embedded in the scanned project files.

swiftui-specialist