The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets and interpolates their content into prompts, creating an attack surface for indirect prompt injection.
Ingestion points: Data is loaded from JSON files (e.g., train.json, test.json) containing lists of strings under feature keys as specified in the Dataset Format section of SKILL.md.
Boundary markers: The prompt templates in references/config_template.yaml (e.g., observations, inference) lack explicit delimiters or instructions for the model to ignore potential commands embedded within the interpolated data.
Capability inventory: The skill executes CLI tools, shell scripts, and performs network-based inferences via LLM APIs.
Sanitization: No sanitization or escaping of the feature variables is defined in the configuration or instructions before they are injected into templates using ${variable} syntax.
[COMMAND_EXECUTION]: The skill instructions involve executing local shell scripts and CLI tools for environment setup and data processing.
Evidence: The skill specifies the execution of bash ./modules/setup_grobid.sh, bash ./modules/run_grobid.sh, and hypogenic_generation / hypogenic_inference CLI commands.
[EXTERNAL_DOWNLOADS]: The skill requires downloading datasets and dependencies from external repositories and registries.
Evidence: The documentation includes commands to git clone datasets from github.com/ChicagoHAI repositories and pip install the hypogenic package.

hypogenic