Can a prompt replace Gumloop?

Build a linear workflow tool where an LLM call is a first-class step type, not just data transformation — Gumloop's actual AI-native idea. Use Next.js/TypeScript with Postgres, and a model key read from ANTHROPIC_API_KEY or OPENAI_API_KEY. Implement three node types usable in sequence: a trigger (webhook or manual run), a transform step (a small explicit JS function), and an AI step (send the current data plus a configured prompt template to the model, use its response as the step's output). Let the user define a workflow as an ordered list of these node types, a config file or a simple form-based builder, not a full drag-and-drop canvas, and run it end-to-end on trigger. Show each step's output before the next runs, both live during a run and in a stored run history, so an AI step's output can be inspected rather than trusted blindly. Store every run in Postgres with per-step status. Do not build a true visual drag-and-drop canvas, a library of pre-built connector nodes for specific services, or team-shared libraries — those are out of scope; workflows here are configured, not visually dragged into place. Requires an Anthropic or OpenAI API key, plus hosting and a database.