The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill includes explicit security guardrails for handling untrusted data. It instructs the agent to treat instructions embedded in untrusted documents or tool outputs strictly as data and explicitly states that such content cannot authorize actions or override safety constraints.
[DYNAMIC_EXECUTION]: The skill implements a conditional loading mechanism where it reads the project's configuration files (e.g., CLAUDE.md or AGENTS.md) to determine if it should load the references/archmage.md persona. This is a localized, safe use of dynamic instruction loading, as it targets a specific file within the skill's own package and is restricted to the active project root.
[PROMPT_INJECTION]: While the skill instructions emphasize honoring user decisions, they maintain a clear boundary by requiring those decisions to remain 'within the host's constraints.' This preserves the underlying safety layer of the AI model while allowing for stylistic flexibility.
[SAFE]: The skill does not request or utilize any dangerous tools, perform network operations, or access sensitive system files. Its primary function is content transformation (prose rewriting) based on a strictly defined set of stylistic 'banned constructions.'

dispel