The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted academic manuscripts, which is an inherent attack surface for indirect prompt injection. However, the risk is comprehensively mitigated by the 'Sprint Contract Protocol' and 'Iron Rules'.
Ingestion points: Manuscripts are ingested in Phase 1 (Panel Review) of the workflow.
Boundary markers: The skill uses explicit XML-style tags (<phase1_output>) and 'Iron Rules' to instruct agents to ignore embedded commands in untrusted materials.
Capability inventory: The agents have no direct network or arbitrary command execution capabilities; they interact with local validation scripts (e.g., scripts/check_panel_synthesis.py) and an orchestrated cross-model verification service.
Sanitization: The architecture employs a two-phase verification where agents must define and acknowledge their scoring triggers in a 'paper-blind' phase before being granted access to the manuscript, preventing adversarial content from altering the evaluation framework.
[DATA_EXFILTRATION]: The skill features an optional 'Cross-Model Reviewer Track' for verifying findings. This process is designed with security best practices, including an explicit user consent gate and strict data-minimization protocols before manuscripts are shared with external model providers.
[COMMAND_EXECUTION]: The orchestration logic refers to internal Python scripts (e.g., scripts/check_pipeline_integrity.py) used for system integrity and consistency auditing. These tools are legitimate vendor resources used for process validation.

academic-paper-reviewer