The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches research paper data and metadata from academic repositories including arXiv (export.arxiv.org) and Crossref (api.crossref.org). These are well-known and trusted services for scientific research.
[COMMAND_EXECUTION]: The freshness protocol in instructions/freshness.md uses the system date command (date -u +%Y-%m-%d) to anchor research findings to the current date. This is a benign and standard practice for ensuring temporal accuracy.
[DYNAMIC_EXECUTION]: The script toolchains/read_any_file.py utilizes the __import__ function to conditionally load optional third-party libraries (e.g., pdfplumber, pandas, beautifulsoup4). This is a standard Python pattern for feature detection and dependency management without requiring all libraries at runtime.
[PROMPT_INJECTION]: The skill processes untrusted external data (research papers) as its primary input. To mitigate indirect prompt injection risks, the skill uses specific delimiters (<<< ... >>>) in its parsing prompts (e.g., prompts/parse-paper.md) and enforces strict constraints requiring the model to adhere exclusively to the extracted text.
Ingestion points: Reads files and URL content via toolchains/read_any_file.py and toolchains/fetch_paper.py.
Boundary markers: Uses <<< ... >>> delimiters in all data-parsing prompts.
Capability inventory: Includes filesystem read/write for caching, network access for fetching papers, and basic shell command execution for the system clock.
Sanitization: Prompt instructions (e.g., prompts/plain-english.md) mandate strict adherence to source text and prohibit the fabrication of facts.

read-research-paper