The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied propositions and arguments which constitutes an attack surface for indirect prompt injection. However, the skill implements robust logical guardrails by requiring the agent to 'rewrite slogans, emotions, or value statements into verifiable propositions' and to construct 'Steelman' (strongest possible) versions of all arguments. These instructions act as a prompt-level sanitization layer that normalizes untrusted input before it influences the agent's logic.\n- [NO_CODE]: This skill consists entirely of Markdown-based guidance, templates, and protocols. There are no Python scripts, Node.js modules, or shell scripts included in the package, which eliminates the risk of remote code execution, binary exploitation, or malicious system-level operations.\n- [SAFE]: All external references are limited to standard documentation links (GitHub) and sibling directory references within the local file system (e.g., ../crossframe/). No network exfiltration, hardcoded credentials, or obfuscated content were found during the analysis.

crossframe-debate