The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted external data (research task descriptions, paper abstracts, and codebase context) through interpolation in prompts located in references/ideation-prompts.md. This architecture presents an indirect prompt injection surface.
Ingestion points: The variables {task_description}, {code}, and {idea} in references/ideation-prompts.md ingest data from the user or previously generated outputs.
Boundary markers: The prompts use XML-style tags (e.g., <experiment.py>) and triple-quote delimiters (""") to separate untrusted content from instructions, which provides some structural protection against accidental instruction following.
Capability inventory: The skill has the capability to perform network requests (via scripts/novelty_check.py) and write execution results to local files.
Sanitization: There is no explicit evidence of input validation, escaping, or filtering of the external content before it is interpolated into the agent context.
[EXTERNAL_DOWNLOADS]: The scripts/novelty_check.py script performs network requests to the Semantic Scholar API (api.semanticscholar.org) to fetch literature metadata and abstracts. While the tool downloads external content at runtime, Semantic Scholar is a well-known service and the script uses standard library functions with secure URL encoding for its parameters.

idea-generation