The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The research pipeline ingests untrusted data from external sources, such as arXiv papers and GitHub repositories. The ingested content is used to inform idea discovery and code generation without explicit sanitization or boundary markers (e.g., delimiters) to prevent potential indirect prompt injection.
Ingestion points: Fetches research papers and metadata from arXiv; clones external repositories from GitHub.
Boundary markers: No specific delimiters or safety instructions are defined for processing external text.
Capability inventory: Extensive use of Bash for command execution and file management; autonomous code generation and modification.
Sanitization: Lacks specific filters or validation for data retrieved from external sources before it influences the agent's logic.
[DYNAMIC_EXECUTION]: The skill is designed to autonomously implement, review, and execute Python experiments and fix cycles. This includes generating new code and running it to verify hypotheses and address reviewer comments.
[COMMAND_EXECUTION]: The skill utilizes shell commands and local Python scripts (run_state.py, iteration_log.py, verify_paper_audits.sh) to manage the research lifecycle, track progress, and validate artifacts.
[EXTERNAL_DOWNLOADS]: The skill fetches documentation and research materials from arXiv and source code from GitHub. These resources originate from well-known services and are integral to the skill's research functionality.

research-pipeline