dev-team

Use the /build command, which plans, implements, tests, and documents a complete vertical slice covering both frontend and backend. For screens it includes screenshots at real device sizes and passing end-to-end tests before handoff.

Use the /debug command, which gathers evidence first (logs, repro steps, stack traces, recent diffs) and confirms the root cause before patching. This avoids hiding bugs behind plausible-looking but unconfirmed fixes.

Defaults include Next.js for web, Expo for mobile, FastAPI or Next.js for backends, Postgres for data, and self-hosted Docker infrastructure. These are defaults for new decisions only, never a mandate to migrate existing working stacks.

Yes, the ai-engineering reference covers defense-in-depth against prompt injection, least-privilege tool scoping, output validation, and access control before retrieval in RAG systems. Model output is always treated as untrusted input.

Fine-tune when the task needs behavior a prompt cannot reliably produce, such as narrow high-volume classification or domain-specific style. Prompt or use RAG when the task benefits from frontier model reasoning and volume does not justify training infrastructure.

It requires an environment where commands, tests, and browsers can actually run, since it never claims unverified work as done. When credentials or test environments are missing, it states what was not verified rather than working around the gap.