The Agent Skills Directory

[SAFE]: The skill implements a strictly offline environment using Python scripts that rely only on the standard library. These scripts include robust safety checks, such as rejecting symbolic links and capping file sizes and JSON nesting depths to prevent resource exhaustion attacks.
[SAFE]: Confidentiality is maintained through explicit instructions forbidding the transmission of sensitive or unpublished data to external services. The included tools are designed to provide audit reports without echoing or exfiltrating the analyzed manuscript content.
[SAFE]: The workflow incorporates a cryptographic verification mechanism where human-verified claims are tied to SHA-256 hashes of specific manuscript lines, ensuring that any changes to scientific prose are detected during the audit phase.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied manuscript files and metadata, which constitutes an ingestion surface for indirect prompt injection. This risk is effectively mitigated by the skill's design, which uses narrow, local tools with no capabilities for arbitrary code execution or outbound network connectivity.

scientific-writing