The Agent Skills Directory

[SAFE]: All nine files in the skill consist solely of markdown instructions and static reference data. There is no evidence of executable scripts, binaries, network exfiltration, or unauthorized file access.
[PROMPT_INJECTION]: The skill processes untrusted user-supplied manuscript data, which creates a surface for indirect prompt injection. However, this risk is mitigated as the skill lacks any tools (such as network access or shell execution) that could be exploited.
Ingestion points: User-provided manuscripts, abstracts, and author notes enter the agent context through SKILL.md and references/reviewer-workflow.md.
Boundary markers: The agent is instructed to use explicit markers like AUTHOR_INPUT_NEEDED and Assessment boundary to delimit analyzed content.
Capability inventory: A manual audit confirms zero capabilities beyond text generation; no subprocesses, environment variables, or network protocols are utilized.
Sanitization: While no explicit input sanitization is performed, the lack of dangerous capabilities prevents any significant impact from malicious input.

nature-reviewer