The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that explicitly override standard AI conversational behavior and safety guardrails regarding human-in-the-loop confirmation. It directs the agent to 'Don't ask' and 'Never say the whys out loud as questions,' prioritizing automated action over clarification.
[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface where malicious user instructions could be executed without prior review due to the 'action-first' policy.
Ingestion points: The skill triggers on user requests starting with 'I need a...' or 'I need this to...' as defined in SKILL.md.
Boundary markers: There are no instructions to sanitize or separate user-provided code/content from the execution context.
Capability inventory: The skill explicitly enables file writing ('written to a file they can open') and command/script execution ('run it. Don't ask') in SKILL.md.
Sanitization: The skill lacks any instructions to validate or sanitize the 'photo' (the generated script or output) before execution.
[DYNAMIC_EXECUTION]: The skill mandates the generation and execution of code at runtime based on user prompts. The instructions 'run it before you show it' for any script or command created from user words constitute a dynamic execution pattern.
[COMMAND_EXECUTION]: SKILL.md instructs the agent to 'run it' whenever a user requests a script or command, which involves executing subprocesses or shell commands directly based on potentially untrusted input.

shutter