The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct users to grant broad shell permissions to agents by adding Bash(*) to the allowed tools in .claude/settings.json during the setup phase.
[PRIVILEGE_ESCALATION]: The workflow recommends configuring the 'Driver' agents with unrestricted file editing and shell access, which represents a high-privilege tool configuration within the development environment.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection due to its core operational model.
Ingestion points: Agents are designed to read and process project source code, documentation (PROJECT.md, glossary.md), and user-supplied product vision statements.
Boundary markers: The provided templates and protocols do not define specific boundary markers or 'ignore' instructions to prevent agents from following malicious prompts embedded in ingested data.
Capability inventory: The 'Driver' agents possess Edit, Write, and Bash(*) capabilities, enabling arbitrary file and system manipulation.
Sanitization: The skill lacks mechanisms for sanitizing or validating external content before it is interpolated into agent contexts.
[DYNAMIC_EXECUTION]: The skill dynamically generates agent definition files (.claude/agents/*.md) and coordinator instructions (.team/coordinator-instructions.md) based on project metadata, which are subsequently loaded to define the behavior and instructions for subagents.

ensemble-team