How the AI team actually builds a full-stack app

The important part of Shipd's AI team is not simply having five AI specialists. It's how the work is passed from one stage to the next.

A single model in a chat box holds one big blurry context and does everything inside it. The team works as a controlled pipeline instead: your requirement goes to product planning, then design direction, then the design system, then the backend and API, then the frontend, and out the other end comes a live preview. Each specialist works with its own context, its own responsibility, and its own output. The output of one stage becomes the input of the next.

A pipeline, not a chat box

The product stage produces the Product Brief and user stories. The design stages turn that into a design direction and then a full design system. The backend stage creates the data model and REST API, and the frontend stage builds against that actual API rather than guessing what the backend should look like.

Which means context accumulates as the build moves right. By the time Sam writes the first component, he's holding Maya's stories, Ren's design system, and Devin's real endpoints.

Every hand-off is gated by you

Between the stages sits a human approval gate. When a specialist finishes, their output stops at the gate: you read the brief, the direction, the design system, or the API, and you approve it or send it back with a note before the next specialist continues.

You're not watching a progress bar and hoping. Every artifact is inspectable before it becomes the foundation for the next stage, and steering early is much cheaper than arguing with a finished app.

Which model runs which specialist

The two coding specialists, Devin and Sam, run on Claude Opus 5 on paid builds, with free builds running on a lighter model from the same Claude family. Both coding roles use the same model deliberately, so the backend and frontend work from the same API contract without interpreting the specification differently.

We don't pin Maya, Theo, and Ren to one named model. The planning and design stages run on the Claude family too, tuned per stage, and the mix shifts as models improve. The selection principle stays constant: the most capable model goes where it earns its cost, which is code generation and the coordination between the two engineers. We wrote about that move in the AI team now ships the backend too.

Devin writes the contract. Sam builds against it.

The engineering half isn't one large generation either. Devin goes first and produces the backend contract: the data model, the API endpoints, and the response structure, then the backend itself. Sam receives that contract and builds the frontend against it.

That ordering is the difference between a frontend wired to a real API and a frontend stitched to a guess.

Feature by feature, checked before continuing

Sam doesn't generate the whole application in one pass. He builds it feature by feature, and each feature is checked in the live preview before the team moves to the next one. If a check fails, that feature gets fixed before anything new is built on top of it.

So the actual shape of the process is understand, plan, design, define the API, build, check, continue. Not prompt, one model, generate the entire app. That separation is the core idea behind Shipd's AI team.

Who the AI team is actually for

Developers. The team moves through product, design, backend, and frontend work without you manually coordinating every part. You still read the outputs like a reviewer, because that's what the gates are for.

Founders. If you know what product you want but don't want to translate every idea into technical instructions, you describe the outcome and participate in the decisions that matter: the scope, the direction, the gates.

Agencies. A client gives you a requirement. Instead of translating it into separate product, design, and development tasks by hand, the team moves through those stages while you review the work at each gate. That's exactly the workflow our agency programme is built around: a PM scoping the work, a director setting the art direction, a designer laying out screens, and engineers building both halves.

Small product teams. Most small teams don't have a dedicated person for every role. An AI team fills some of that gap without pretending AI has replaced human engineering or product judgment. The judgment stays with you; the gates make sure of it.

If you want the tour of the five specialists themselves, that's meet your AI team. If you want to see the pipeline run, brief them once and watch.

Send the team to work →