The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided natural language descriptions to generate application code. This architectural pattern is inherently susceptible to indirect prompt injection if the user input contains malicious instructions. However, the skill implements a robust mitigation strategy through a mandatory 'Phase 1.5' Plan Verification checkpoint.
Ingestion points: User requests defining project features and logic (processed by SKILL.md and project-detection.md).
Boundary markers: The skill requires the creation and verification of a PLAN.md file in the project root before any specialist agents are permitted to proceed with development (documented in agent-coordination.md).
Capability inventory: The orchestrator utilizes a powerful toolset including Read, Write, Edit, Glob, Grep, Bash, and Agent (SKILL.md).
Sanitization: Explicit sanitization is not defined, but the multi-step agent pipeline and planning phase serve as structural guardrails to ensure user intent is validated before code execution.
[EXTERNAL_DOWNLOADS]: The skill templates instruct the agent to use official package managers and CLI tools (npx, npm, pip, flutter) to initialize projects. These downloads target well-known, trusted framework providers including Next.js (Vercel), Nuxt, Vite, Expo, and FastAPI. The dependency lists in the templates use standard packages from the npm registry and PyPI.
[COMMAND_EXECUTION]: The skill uses the Bash tool for project scaffolding tasks, such as creating directories, initializing repositories, and installing dependencies. This use of shell access is consistent with the primary purpose of an application builder and follows standard developer workflows.

app-builder