The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from untrusted external sources (academic APIs) and local user files, which could theoretically contain malicious instructions designed to influence the agent's behavior during citation processing.
Ingestion points: The skill reads local LaTeX (.tex) and BibTeX (.bib) files via scripts/check_bib.py and fetches external content from DBLP, Semantic Scholar, arXiv, and Crossref APIs via scripts/cite_lookup.py.
Boundary markers: The skill implements a verification process requiring all citations to be logged in citations_log.md and mandates the use of placeholders for unverified content, which acts as a manual check against potential injection.
Capability inventory: The skill has the ability to perform network requests (urllib.request.urlopen), write to the filesystem (creating the .icml/ directory and logs), and is intended to be used alongside shell command execution.
Sanitization: External data is parsed using structured formats like JSON and XML, though the agent interprets the semantic content of titles and abstracts to summarize them for the user.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch citation metadata and BibTeX records from well-known academic services.
Details: scripts/cite_lookup.py connects to dblp.org, api.semanticscholar.org, export.arxiv.org, and doi.org. These are established scientific data providers appropriate for the skill's primary purpose.

icml-cite