The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided feature requirements and design inputs to generate implementation plans that drive subagents. While this creates a theoretical surface for indirect injection, the risk is mitigated by built-in design constraints.
Ingestion points: User input via the /build-feature command and design details gathered during the 'grill-me' interview phase, which are used to populate design specs and implementation DAGs (Directed Acyclic Graphs) in SKILL.md.
Boundary markers: The workflow uses structured JSON files (lego-plan.json) and specific markdown files for planning, providing clear structural separation between instructions and data.
Capability inventory: The skill possesses capabilities for file system operations (creating and opening specs/plans), subagent spawning for task execution, and coordination of multiple specialized quality-control skills.
Sanitization: Security is enforced through mandatory human-in-the-loop 'Hard Gates'. The skill strictly forbids proceeding to implementation until the user explicitly approves the generated design and implementation plan, preventing the autonomous execution of potentially poisoned instructions.
[SAFE]: The skill follows security best practices by maintaining transparency and user control throughout the automation process.
Transparency: It enforces the continuous rendering of a Mermaid-based DAG in the chat, ensuring the user can monitor every state change and node status (Hard Gate 8 and 9).
Least Privilege: It restricts implementation subagents from accessing the planning or interview tools, limiting their scope to specific files and quality standards (Hard Gate 6).
Controlled Environment: File operations are restricted to specific project documentation paths (docs/superpowers/), and the skill explicitly avoids committing unapproved artifacts to the repository.

build-feature