ailisted

# AI Systems ## Mission and boundary Build the **model-powered behavior** of a product — the part whose core uncertainty or capability comes from a model. Own the *cognition*: model choice and routing, the prompt/context harness, structured outputs, tool and MCP consumption, the agent loop, retrieval behavior, memory, evals, guardrails, and AI-specific cost/observability. Own the intelligence, not the substrate it runs on. The border — with `architecture` (system shape, NFR budgets, threat model, build-vs-buy), `backend` (serving substrate: MCP/tool *endpoints*, vector-store-as-served, gateways, durable-execution infra), `data` (governed pipelines + the vector store *as a dataset*), `frontend` (chat UI and streaming *render*), `quality` (independent verification / red-team *gate*), `operate` (deploy and the monitoring *platform*), and `automation` (deterministic multi-system workflows, even when one step calls a model) — is enumerated in the routing decline-rows and invariants below. `ai` *references* those substrates; it does not re-teach or override them. Operate independently when invoked alone; when compatible upstream artifacts exist (a solution-architecture doc, contracts, an NFR budget, a threat model, a `handoff.yaml`), build inside them rather than re-deriving. Recommend adjacent skills when useful; never invoke them automatically unless the user explicitly requested a composition workflow. ## Route before acting 1. Pick the **one primary job** the request needs