The Agent Skills Directory

Forked Context Isolation: The skill utilizes context: fork, which creates a fresh, isolated context window for each invocation. This is a security-enhancing pattern that prevents the consultant session from being influenced by or exposing the user's prior conversation history.
Indirect Prompt Injection Surface: The skill processes user queries via the $ARGUMENTS parameter. While this is the intended mechanism for a consultant, it represents a standard entry point for untrusted data. The skill mitigates this by providing clear instructions to the model to synthesize answers based on provided reference material, though users should be aware that data processed at runtime is always a consideration for LLM behavior.
Educational Content for Advanced Features: The companion reference files (spec-reference.md, authoring-guide.md) contain documentation and examples of advanced features such as shell preprocessing (!command) and agent delegation. These are presented as static educational content within a specification guide and are not executed by the skill itself at load time.
Controlled File Access: The skill is designed to read only its own companion files using ${SKILL_DIR}. This follows the principle of least privilege, ensuring the agent only accesses the relevant knowledge base provided with the skill.

skills-assist