The Agent Skills Directory

[SAFE]: The skill implements a structured orchestration framework with strong safety guardrails. It explicitly restricts its own scope to pre-approved specifications and non-destructive validation. It requires returning control to a human agent whenever changes involve external systems, destructive operations, significant costs, or conflicting authoritative materials.
[SAFE]: It employs a clear privilege separation and review model. The skill mandates separate sessions for implementation and review tasks, ensuring that reviewers do not inherit the assumptions or conclusions of implementers. It also includes explicit risk-tiering (T1/T2) to adjust model intensity based on the complexity and security sensitivity of the task.
[SAFE]: Data handling practices are secure. Instead of passing sensitive content directly, the skill uses 'context pointers' (file paths or Git commit hashes), promoting the use of existing environment-controlled access methods rather than creating new data exposure vectors.
[SAFE]: The skill includes built-in persistence for auditing purposes, recording task metadata and outcomes to local JSONL files in a designated scratch directory (.scratch/afk-runs/). This facilitates transparency and post-execution review without attempting to maintain unauthorized persistence in system-critical locations.

afk-impl