The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from user-provided PDF documents to extract bibliographic information, creating a surface for indirect prompt injection. Maliciously crafted text in a PDF's reference list could attempt to influence the agent's logic during parsing or metadata enrichment phases. \n
Ingestion points: The split-pdf tool is used to read user-provided PDF files in Phase 1.1.\n
Boundary markers: The skill includes a mandatory manual confirmation step in Phase 1.2, where the user must verify the parsed reference skeleton before enrichment proceeds.\n
Capability inventory: The agent has access to Bash(uv*), WebSearch, WebFetch, and file writing capabilities.\n
Sanitization: Outputs are validated via the bib-validate skill in Phase 4.2 and a local verification script in Phase 4.3.- [COMMAND_EXECUTION]: Extracted metadata from the PDF, such as titles and DOIs, is passed as arguments to CLI tools. In Phase 2.2, titles are interpolated into commands like scholarly scholarly-search "<title>". While the use of quotes provides some protection, there is an inherent risk of command injection if the extraction logic fails to sanitize special characters within the PDF text.- [EXTERNAL_DOWNLOADS]: The skill utilizes WebSearch and the scholarly CLI to fetch data from external academic databases, including OpenAlex, Semantic Scholar, Scopus, and Web of Science. These are well-known services required for the metadata enrichment functionality of the skill.

bib-parse