The Agent Skills Directory

[SAFE]: The skill provides a structured approach to designing AI agent personas, focusing on identity, voice, and tone. It does not contain any prompt injection or bypass instructions.\n- [REMOTE_CODE_EXECUTION]: The skill uses a local Python script (scripts/parse_agent_yaml.py) to parse agent configuration files in brownfield mode. The script is read-only, uses standard regex for extraction, and does not execute remote code or download external dependencies.\n- [DATA_EXFILTRATION]: No evidence of data exfiltration was found. The skill reads local files to perform its analysis and writes output to a local persona/ directory. No network tools like curl or wget are used to communicate with untrusted domains.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text (brand guides, organization URLs) and existing agent files. While this represents a potential attack surface for indirect prompt injection, the risk is inherent to the skill's primary purpose as a design tool and is managed through the use of structured templates and encoding guidelines. This is assessed as safe within the context of its intended use-case.\n- [OBFUSCATION]: A thorough review of all skill files, including scripts, templates, and reference guides, revealed no hidden characters, Base64-encoded instructions, or other obfuscation techniques.\n- [CREDENTIALS_SAFE]: No hardcoded API keys, secrets, or sensitive file paths (e.g., .ssh, .aws/credentials) are accessed or exposed by the skill.

agentforce-persona-generate