The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses an indirect injection surface by reading external SKILL.md files from the local file system and hydrating their content as active instructions.
Ingestion points: The view_file tool is used to read SKILL.md files from paths like ~/github/ and /usr/local/google/home/kevmoo/....
Boundary markers: None. The skill instructions do not specify any delimiters or warnings to ignore embedded instructions within the hydrated content.
Capability inventory: The agent is granted the ability to read files and perform refactoring based on the content of those files.
Sanitization: None. Content read from the local files is directly incorporated into the agent's execution context.
[COMMAND_EXECUTION]: The skill implements a 'Hydration & Execution' protocol that dynamically loads instructions from computed file paths based on user intent. This pattern of dynamic instruction loading from the file system is a high-privilege behavior that relies on the integrity of local checkouts.
[DATA_EXPOSURE]: The skill hardcodes absolute local file paths belonging to a specific user directory (/usr/local/google/home/kevmoo/...). While these represent vendor-owned resources for the author 'kevmoo', they expose the specific directory structure of the host environment.

dart-toolkit