The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements significant safety constraints for shell usage, explicitly prohibiting dangerous commands (e.g., 'sudo', 'rm -rf /', 'chmod 0777') and enforcing workspace isolation by forbidding the use of absolute paths outside the designated project directory.
[EXTERNAL_DOWNLOADS]: It restricts remote resource fetching by prohibiting 'curl | sh' and manual package managers ('pip install'), instead promoting the use of an authorized internal skill whitelist for task-specific capabilities.
[CREDENTIALS_UNSAFE]: Includes a mandatory safety protocol to redact secrets and sensitive tokens from response history used during automated retry cycles, preventing accidental credential exposure.
[PROMPT_INJECTION]: Addresses the surface for indirect prompt injection by defining strict output contracts and providing instructions that emphasize context isolation and the rejection of adversarial role-play or stance-mixing.

minimax-prompting