Product Design with Agentic AI

You cannot direct an agent in software you cannot operate yourself — if you do not know what "scale", "extrude", or "the origin" mean, you cannot tell the agent what went wrong or even see that it did. Before the first prompt, you need the moves.

Ask an agent to "make a chair" and it returns something chair-shaped, floating above the floor, three centimetres tall, with its origin marooned at the world centre — the failure isn't the geometry, it's every constraint a human designer holds without noticing.

Hand the agent a reference photo and it will confidently describe a different object than the one in the image — agents narrate what they expect a "lamp" to be, not what they see, so proportion drifts toward the average of their training data.

The agent's model looks fake for a reason you can name in one word — edges. Real objects have no perfectly sharp edges; the agent leaves every edge razor-sharp because nothing in the prompt told it that perfection is the tell.

Ask for "five variations" and you get either five clones with a different colour or five unrelated objects — the agent has no concept of which features are the brand and which are the variable, so it randomises the wrong axis.

The agent will always tell you it is finished — "done" is its favourite word — so the one judgment it can never make for you is the only one that matters at this stage: is this model actually ready to become a game asset?