The Agent Skills Directory

[PROMPT_INJECTION]: The skill's system instructions in prompts/persona_system.md direct the model to ignore standard AI assistant constraints, such as helpfulness and neutrality, to maintain the fidelity of the persona simulation. These instructions utilize role-playing patterns that resemble techniques used to bypass AI safety filters.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface by processing biographical narratives from external sources and interpolating them into context prompts. This could allow instructions embedded within the narrative data to influence agent behavior.
Ingestion points: Narrative fields from datasets listed in data/manifest.json (e.g., nemotron_usa).
Boundary markers: The implementation does not use explicit delimiters to separate the untrusted narrative content from the agent's instructions.
Capability inventory: The skill can perform network requests to LLM provider endpoints and access local configuration files.
Sanitization: No filtering or validation is performed on the ingested narrative data.
[EXTERNAL_DOWNLOADS]: The skill provides automated methods for downloading research-grade datasets from Hugging Face and Zenodo as part of its data acquisition workflow.
[SAFE]: The skill manages credentials securely by instructing users to store API keys in an external config.json file and using environment variables to locate the configuration, which prevents secrets from being hardcoded or included in the repository.

persona-sim