The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that were flagged by automated detectors as attempts to override system rules (e.g., in agents/devils_advocate_reviewer_agent.md). However, analysis confirms these are defensive patterns. The agents are explicitly instructed to ignore potential injection attempts found within untrusted manuscripts by treating delimited content strictly as data rather than instructions.- [SAFE]: The skill is designed to process external academic manuscripts, creating an attack surface for indirect prompt injection. It mitigates this risk effectively:
Ingestion points: Untrusted paper content is ingested in SKILL.md and processed by the reviewer agents in Phase 1 and 2.
Boundary markers: The skill uses <phase1_output> data delimiters to separate prior agent commitments from manuscript content during processing.
Capability inventory: Reviewer agents are limited to generating text-based reports; all file synthesis is handled by a separate synthesizer agent, and reviewers are strictly prohibited from modifying the manuscript.
Sanitization: Instructions explicitly state that embedded directives in manuscripts must not alter agent identity, routing, or workflow constraints.- [COMMAND_EXECUTION]: The skill references local scripts (scripts/check_sprint_contract.py and scripts/check_pipeline_integrity.py) used for internal validation of the multi-agent review protocol. These scripts are part of the skill's infrastructure and do not involve unauthorized execution or remote downloads.

academic-paper-reviewer