The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/lint-goal-prompt.py) to validate the syntax and safety of goal prompts. This execution is performed on local files as part of the skill's validation workflow.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data in the form of goal prompts or files to be linted, creating a vulnerability surface. Mandatory Evidence Chain: (1) Ingestion points: Data enters the context via file reading in scripts/lint-goal-prompt.py. (2) Boundary markers: The linter script explicitly checks for and rejects vague instructions and placeholders (e.g., 'until it feels good'). (3) Capability inventory: The skill uses subprocess calls to execute the local linter script. (4) Sanitization: The linter script sanitizes input by verifying it against required labels and safety constraints before adoption.
[SAFE]: The skill implements a 'Safety Gate' mechanism that explicitly blocks autonomous execution of destructive, irreversible, or sensitive operations (e.g., authentication, billing, production data changes) without human approval. This design significantly reduces the risk of unauthorized actions or agent drift.

goal-gate