The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses persona-adoption instructions such as "Become the specified character" and "Speak in first person" to define the agent's conversational behavior. These instructions are focused on psychological and stylistic realism within a role-play context and do not attempt to override safety filters or disregard system-level constraints.
[DATA_EXFILTRATION]: While the skill instructions refer to external data sources like "style files" and "character state files," it does not contain any commands to exfiltrate data, access sensitive local directories (e.g., .ssh, .aws), or make unauthorized network connections.
[REMOTE_CODE_EXECUTION]: There are no patterns of remote script execution, package installation, or shell command invocation. The skill operates entirely within the conversational context of the LLM.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data (character sketches and style files) which could potentially contain malicious instructions. However, the skill lacks any execution capabilities or tools, meaning it cannot perform actions outside of generating text, which mitigates the risk.

character-sim