The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes manuscript Markdown, intake JSON, and claim-evidence CSV files. This attack surface is mitigated by using deterministic local tools that perform structural and consistency checks rather than executing embedded instructions. The workflow requires a human reviewer to verify and finalize all generated scaffolds, and the tools are designed to emit only identifiers and counts to the agent rather than echoing manuscript prose.
[SAFE]: The skill implements defensive coding practices in scripts/_common.py, such as the rejection of symlinks to prevent path traversal, strict input size limits (4MB), and atomic file writes with restricted owner-only permissions (0o600).
[SAFE]: The skill includes a detailed security_validation.md record documenting the removal of previous high-risk components and the implementation of AST-based checks to ensure no network or dynamic code execution libraries are used.
[DATA_EXFILTRATION]: Instructions in SKILL.md and the compatibility statement explicitly prohibit network calls, external service use, and the transmission of confidential manuscript data to external models. The bundled Python tools use only the standard library and do not include any network-capable modules.
[SAFE]: External references and documentation pointers target well-known and trusted scientific domains (e.g., arXiv, EQUATOR Network, CONSORT, and official journal policy pages), which is appropriate for the skill's intended research use case.

peer-review