The Agent Skills Directory

[SAFE]: The skill interacts with established academic services including Semantic Scholar, OpenAlex, Crossref, BASE, and the German National Library (DNB). These are well-known, reputable sources for scholarly metadata and publications.
[CREDENTIALS_SAFE]: Instructions correctly direct the agent to retrieve API keys from environment variables (e.g., SEMANTIC_SCHOLAR_API_KEY) and explicitly warn against hardcoding or committing secrets.
[COMMAND_EXECUTION]: The skill uses curl to fetch bibliographic data and download Open Access PDFs to a temporary directory (/tmp/paper.pdf). Text extraction is performed using local tools like pdftotext or standard Python libraries. This represents a legitimate and scoped use of shell commands for data processing rather than arbitrary execution.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (abstracts and full-text PDFs) which constitutes an ingestion surface for untrusted content.
Ingestion points: Data enters via curl requests to academic APIs and PDF downloads.
Boundary markers: The instructions mandate strict output formatting and the use of verbatim quotes, which helps the agent distinguish between source content and its own reasoning.
Capability inventory: The skill has access to shell commands (curl, pdftotext) and file writing (to /tmp).
Sanitization: The skill emphasizes facts, quotes, and sources, requiring the agent to "bind claims to evidence" and "label clearly what the source states vs. what you infer."
Note: While a malicious PDF could attempt to inject instructions, the skill's rigorous synthesis workflow and focus on structured evidence extraction provide effective mitigation.

academic-research