The Agent Skills Directory

[SAFE]: The skill is a management component designed to automate task tracking and workflow execution. It uses standardized tools like spawn_agent for delegation and TaskUpdate for progress tracking.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface for indirect prompt injection because it executes commands and sequences parsed from local project files (workflows.json, CLAUDE.md). If these configuration files were modified by an adversary, the orchestrator would execute the modified instructions.
Ingestion points: Workflow definitions are read from .claude/workflows.json and context catalogs in CLAUDE.md and AGENTS.md.
Boundary markers: The protocol implements a 3-tier lookup strategy to ensure it only activates defined workflow IDs from the catalog.
Capability inventory: The skill utilizes file reading (grep), task state management (TaskUpdate), and sub-agent spawning (spawn_agent) to fulfill its orchestration role.
Sanitization: The protocol relies on the integrity of the local repository configuration and does not include explicit sanitization of the workflow step IDs before execution.

start-workflow