The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill framework processes external workloads and evaluation feedback, which represents a surface for indirect prompt injection. This risk is managed through defined safety constraints and authority requirements.
Ingestion points: External data is ingested through 'representative workloads' and 'judged outcomes' (SKILL.md, references/judged-outcomes.md).
Boundary markers: The instructions mandate 'hard constraints' and 'acceptance constraints' (including security) before starting, along with blinding labels during evaluation.
Capability inventory: The skill utilizes tools like alaga, ko-skill, and seda-pr to finalize and publish changes based on trial results.
Sanitization: Actions like installation, external disclosure, and destructive changes are subject to their own authority requirements.
[SAFE]: No malicious patterns or security vulnerabilities were identified. The instructions emphasize best practices for controlled experimentation, including budget limits and the rejection of false improvements.

optimize