The Agent Skills Directory

[SAFE]: The skill consists of markdown-based instructions and YAML configuration. It does not contain scripts, shell commands, or network operations that could be used for malicious purposes.
[SAFE]: The instructions incorporate defensive measures, such as a 'no-fabrication rule' and a 'Post-edit eval' process, to prevent the agent from inventing facts or being manipulated by external content during the review process.
[SAFE]: Evaluation of the skill's attack surface for indirect prompt injection: (1) Ingestion points: The skill is designed to process user-provided UI code, screenshots, and prose drafts as input for review. (2) Boundary markers: The skill instructions provide explicit constraints, requiring the agent to 'preserve facts' and 'identify existing tokens', which limits the scope of influence from the input data. (3) Capability inventory: The skill is limited to inspection and making suggestions (read/proposal authority) and does not request access to dangerous tools or system-level permissions. (4) Sanitization: The skill defines a formal verification phase (Post-edit eval) to ensure the agent's output is truthful and maintains source-faithful voice, effectively mitigating potential injection risks.

anti-slop-review