The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data in the form of notes, outlines, and source material which are interpolated into generation prompts. While the skill uses '---' delimiters to frame this material, it lacks clear instructions for the model to ignore embedded instructions or commands within the user's input. The mandatory evidence chain is as follows: (1) Ingestion points: 'user_provided_source_material' in SKILL.md and 'source_material' in modules/source-framing.md. (2) Boundary markers: DELIMITERS ('---') are present in both files. (3) Capability inventory: The skill drafts content using the Claude 3 Opus model and executes local shell commands (ls, grep). (4) Sanitization: No sanitization or filtering logic is present in the skill files.
[COMMAND_EXECUTION]: The skill modules, specifically modules/register-selection.md, include bash logic that executes commands such as 'ls', 'grep', 'basename', and 'tail' to manage and list voice profile configurations stored in the user's home directory. While these operations support the skill's primary function, they represent a script-based capability that interacts directly with the file system.

voice-generate