The Agent Skills Directory

Dynamic Skill Generation: The skill is designed to author new instruction sets (SKILL.md files) at runtime. This process involves generating executable content (markdown instructions) based on interactive user input, which the agent then follows in subsequent tasks.
Persistence via Workflow Registration: The skill registers newly created automation into a central routing configuration (smb-router). This acts as a persistence mechanism, allowing the agent to automatically trigger these new behaviors in future sessions based on specific keywords.
Indirect Prompt Injection Surface: The skill uses untrusted user transcripts and descriptions to define the logic of new workflows.
Ingestion points: Task descriptions and conversation transcripts in SKILL.md and reference/capture.md.
Boundary markers: The agent is instructed to confirm the task sequence with the user before finalization.
Capability inventory: The skill uses Read and WebFetch tools, and the authoring process involves file system writes to create the new skill and modify the router.
Sanitization: The skill relies on AI interpretation and human-in-the-loop approval gates for consequential actions (sending, spending, or deleting) rather than automated sanitization.
External Configuration Access: The skill instructions involve reading sensitive user data from relative paths outside the local skill directory, such as a voice profile at ../../../shared/voice-profile.md, which creates a dependency on platform-level shared data.

build-agent