The Agent Skills Directory

[SAFE]: The skill incorporates comprehensive safety guardrails that mandate explicit user approval for all semantic changes, preventing the agent from performing autonomous or hidden modifications to the codebase.
[SAFE]: In accordance with the shared guardrails, the skill instructs the agent to refuse any implementation that would weaken security controls, trust-boundary validation, or data-loss prevention.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted code and documentation as its primary input (ingestion points: target files being patched). It effectively manages the risk of instructions embedded in data through mandatory boundary markers, specifically a phased approval workflow requiring the presentation of unified diffs and explicit approval intent tokens from the user before execution. The skill's capability inventory (file operations and command execution) is strictly gated by these reviews and subsequent verification checks (sanitization).

duck-patch