The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The system retrieves academic metadata and paper information from well-known and reputable services.
The framework interacts with APIs from arXiv, Crossref, OpenAlex, and Semantic Scholar for source verification and discovery.
[COMMAND_EXECUTION]: The framework utilizes standard local utilities to extract information from research documents.
The timeline_extraction_agent is designed to use the pdftotext utility to retrieve publication dates and metadata from PDF files.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and analyzes large volumes of untrusted third-party data, including paper abstracts and full-text documents, which presents a potential injection surface.
Ingestion points: External content enters the system via the bibliography_agent and source_verification_agent during literature searches and source assessments.
Boundary markers: The skill explicitly implements a "canonical:instruction-data-boundary" which mandates that retrieved external content is treated strictly as data and never as executable instructions.
Capability inventory: Individual agents are granted limited toolsets (e.g., Read, Write, Grep) and are strictly constrained by "Phase Boundaries" to prevent unauthorized actions.
Sanitization: The multi-agent orchestration and specialized personas (such as the Devil's Advocate and Ethics Review agents) provide layers of validation and critical oversight to ensure ingested data does not manipulate the final research output.

deep-research