The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from JSON datasets and PDF research papers, which constitutes an indirect prompt injection surface.
Ingestion points: Training, validation, and test datasets in JSON format (referenced in config.yaml), along with PDF files located in the literature/ directory.
Boundary markers: The skill uses structured YAML templates in references/config_template.yaml with defined system and user roles and ${placeholder} variable substitution. No explicit instructions are present to disregard potential instructions embedded in the data.
Capability inventory: Access to Bash (uv, python, git, and package-specific CLI tools), Read, Write, and Edit file operations, and network access for LLM API calls and GitHub cloning.
Sanitization: The implementation uses regex-based extraction (extract_label) to parse LLM outputs; however, there is no evidence of sanitization or filtering applied to the input data before it is interpolated into prompts.
[EXTERNAL_DOWNLOADS]: The skill downloads resources from external repositories to function.
Fetches example datasets from the ChicagoHAI organization on GitHub (https://github.com/ChicagoHAI/HypoGeniC-datasets.git and https://github.com/ChicagoHAI/Hypothesis-agent-datasets.git).
Installs the hypogenic package from the official Python Package Index (PyPI) using uv.
[COMMAND_EXECUTION]: The skill performs several command-line operations to set up and run the hypothesis generation pipeline.
Utilizes uv pip install for dependency management.
Uses git clone to retrieve datasets from GitHub.
Executes local shell scripts (setup_grobid.sh, run_grobid.sh) to initialize the GROBID service for PDF processing.
Runs Python scripts and package-specific CLI entry points (hypogenic_generation, hypogenic_inference) to perform its primary hypothesis testing and generation functions.

alterlab-hypogenic