The Agent Skills Directory

[SAFE]: The skill implements a structured auditing process for project management workflows. It uses a human-in-the-loop model, explicitly stating that it must report findings and obtain approval before performing any 'low-risk' repairs. It does not exhibit malicious patterns such as credential exfiltration, unauthorized network access, or persistence mechanisms.
[PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it ingests untrusted data from project records. * Ingestion points: Reads external 'work records', 'comments', 'receipts', and 'linked implementation artifacts' (SKILL.md). * Boundary markers: The instructions lack explicit delimiters to isolate external content from the agent's core logic. * Capability inventory: The skill can modify backend metadata (e.g., labels, stage states) via a backend reference tool. * Sanitization: No specific sanitization or filtering logic is provided for the ingested data. * Note: The risk is effectively mitigated by the skill's restrictive repair policy and the requirement for explicit user approval for any modifications.

and-sweep