The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (paper abstracts) fetched from external academic APIs such as OpenAlex and Semantic Scholar, and incorporates this data into the AI agent's context for relevance scoring and review synthesis. This creates a surface for indirect prompt injection where adversarial text in an abstract could influence the agent's behavior.
Ingestion points: Literature metadata and abstracts are retrieved via scripts like openalex_search.py and stored in local .jsonl files (e.g., papers.jsonl).
Boundary markers: The ai_scoring_prompt.md uses Markdown headers as delimiters to separate instructions from paper data, providing minimal protection against adversarial content.
Capability inventory: The skill utilizes subprocess to execute system commands including xelatex, bibtex, and pandoc through scripts like compile_latex_with_bibtex.py and convert_latex_to_word.py.
Sanitization: The script build_reference_bib_from_papers.py includes a _sanitize_unicode function to strip Unicode control characters, but no explicit sanitization is performed to detect or block prompt injection patterns in the ingested abstracts.
[COMMAND_EXECUTION]: The pipeline automates the generation and conversion of scholarly documents by executing local system tools via the shell.
Evidence: compile_latex_with_bibtex.py executes xelatex and bibtex to render PDF documents.
Evidence: convert_latex_to_word.py utilizes pandoc to export document drafts into Word format.
[EXTERNAL_DOWNLOADS]: The skill connects to well-known academic metadata providers and price monitoring sources.
Sources: Metadata is fetched from OpenAlex (api.openalex.org), Semantic Scholar (api.semanticscholar.org), and Crossref (api.crossref.org), which are established and trusted services in the research community.
Sources: The pipeline_cost.py script leverages search capabilities to retrieve current AI model pricing from official vendor websites including OpenAI and Anthropic.

systematic-literature-review