The Agent Skills Directory

Command execution for custom model integration: The skill allows users to configure custom LLM execution commands through environment variables such as NOVELTY_LLM_REASONING_LARGE_CMD and NOVELTY_LLM_CLASSIFY_FAST_CMD. These commands are executed via subprocess.run with shell=True in several scripts (run.py, pattern_summary.py, intent.py). While this enables flexibility for integrating different AI backends, it is a potential consideration if the environment configuration is accessible to untrusted actors.
Retrieval and processing of external literature: The skill automatically fetches research paper data from public academic repositories including arXiv, Semantic Scholar, OpenAlex, and OpenReview. This creates a potential surface for indirect prompt injection, where adversarial text embedded within a research paper could attempt to influence the agent's logic. The skill uses detailed system prompts and specific "Hard rules" as mitigation strategies to maintain process integrity.
Automated downloads from established services: The skill's workflow includes downloading paper metadata and full-text sections (via fetch_sections.py) from reputable academic platforms. This behavior is central to the skill's purpose of ensuring research ideas are grounded in existing literature.
Dynamic module verification: During initialization and diagnostics, the skill uses importlib.import_module in run.py to check for the presence of required Python dependencies. This dynamic approach is used to provide clear feedback and setup guidance to the user.

idea-spark