The Agent Skills Directory

[SAFE]: The skill implements a 'Multi-perspective academic paper review' which is a standard scholarly activity. The architecture uses multiple specialized agents (eic_agent, methodology_reviewer_agent, etc.) to process user-provided manuscripts.
[SAFE]: The skill establishes an 'IRON RULE — READ-ONLY CONSTRAINT' (Checkpoint Rule #6 in SKILL.md), explicitly stating that reviewers must not modify the submitted manuscript and only produce separate report documents. This prevents accidental or malicious data modification.
[SAFE]: The skill uses well-defined phase boundaries (v3.9.2) to restrict agent actions, ensuring that evaluation agents cannot write to unauthorized directories or bleed into other stages of the workflow.
[SAFE]: It includes a 'Untrusted Review Materials' rule (Checkpoint Rule #7) that explicitly warns the agent to ignore any embedded instructions inside user-submitted manuscripts that might attempt to alter the skill's routing, tool use, or security constraints.
[SAFE]: All external references and templates (quality_rubrics.md, statistical_reporting_standards.md) are internal to the skill or point to established academic standards (e.g., APA 7.0, Nature, Nature Baselines).
[SAFE]: The 'Devil's Advocate' agent is designed with a specific 'Attack Intensity Preservation Protocol' to prevent conversational bias (sycophancy) during re-reviews, which is a common limitation in LLM agents and is addressed here as a quality feature rather than a security risk.

academic-paper-reviewer