The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data such as paper titles and abstracts from external academic databases (CNKI, Web of Science, ScienceDirect, PubMed) through browser automation. This data is then used as context for the LLM during paper analysis and synthesis phases. Maliciously crafted content within academic paper metadata could attempt to influence the agent's behavior during these processing steps.
Ingestion points: External data extracted from academic databases via the browser tool and stored in papers_raw.json.
Boundary markers: The skill uses delimited markers like <articles>, <collected_literature>, and <detailed_analysis> in its prompt templates to separate ingested content from system instructions.
Capability inventory: The skill uses the browser tool for navigation and data extraction, writes results to the local filesystem (./sessions and ./output folders), and executes Python scripts for data normalization.
Sanitization: Prompts do not contain explicit warnings to ignore or sanitize possible instructions embedded within the ingested paper data.
[PRIVILEGE_ESCALATION]: The skill documentation in SECURITY.md provides configuration advice recommending the use of chmod 777 on the sessions and output directories to resolve permission issues when running within Docker. While provided as documentation for user environment setup and not as executable code within the skill, this recommendation encourages overly permissive access to local folders.
[EXTERNAL_DOWNLOADS]: The skill relies on standard, well-known dependencies for its functionality, specifically requiring the installation of Playwright and its browser binaries (playwright install chromium) to perform automated web searching. It also references and acknowledges several public GitHub repositories used as architectural inspiration.

literature-reviewer-skill