The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a framework for interpolating task-specific data into an assignment template for subagents. This creates a surface where instructions embedded in the task context could influence subagent behavior.
Ingestion points: The file assets/assignment-template.md defines multiple fields (such as 'Objective', 'Read first and context', and 'Owned scope') that ingest context-specific data at runtime.
Boundary markers: Present. The template uses Markdown headers and instructional text to delimit input fields and provide structure.
Capability inventory: The SKILL.md file identifies the use of subagent launchers (native tools, CLI, or plugins) which can perform implementation, research, and inspection tasks across the codebase.
Sanitization: Present. The instructions in SKILL.md and assets/assignment-template.md explicitly warn against including secrets, tokens, or configuration files (e.g., .env) in the assignment context.
[SAFE]: The skill incorporates robust security practices for subagent management, including 'least privilege' permissions, strict working directory boundaries, and a prohibition on recursive delegation where subagents might spawn other agents.
[SAFE]: Mandatory cleanup instructions are provided in SKILL.md, requiring the parent agent to verify completion and ensure that all temporary resources, such as background processes, network sockets, or worktrees, are removed after the task is finished.

use-subagents