The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates a multi-agent workflow where codebase data and user prompts are passed to subagents using the runSubagent function, creating a surface for indirect instructions to influence subtask execution.
Ingestion points: Workflow steps in SKILL.md and patterns in references/patterns.md ingest file content (API routes, data models, component props) to generate code and documentation.
Boundary markers: The skill's 'Verification Protocol' explicitly mandates that context passed to subagents be 'minimal, task-local, and free of hidden expected answers' to prevent unintended cross-task influence.
Capability inventory: The workflow leverages runSubagent for task delegation and mcp_sequentialthi_sequentialthinking for architecture planning, which involves deep filesystem interaction and automated code generation.
Sanitization: A five-step 'Integration Workflow' in SKILL.md requires manual review and validation of subagent outputs. Additionally, a 'Quality Control' checklist specifically includes a check for 'security vulnerabilities' in generated code prior to merging.

subagent-delegation