The Agent Skills Directory

[SAFE]: The skill implements a robust safety framework that enforces multi-stage human approval for all workspace-changing actions, including code edits and task delegation.
[INDIRECT_PROMPT_INJECTION]: While the skill processes user requests, it mitigates risk by requiring explicit confirmation and approval checkpoints before any mutation occurs. 1. Ingestion points: User instructions processed during session startup and request classification in SKILL.md. 2. Boundary markers: Implements mandatory safety gates (Checkpoints 1-4) that act as clear delimiters between analysis and execution. 3. Capability inventory: Instructs the agent on logic edits, command execution, and file operations. 4. Sanitization: Relies on mandatory human-in-the-loop review of diffs and plans as the primary defense against malicious instructions in data.
[COMMAND_EXECUTION]: The skill defines rigorous constraints for command execution, requiring a preflight checklist and explicit user approval intent before any mutating command is run.

duck-policy