“Move the window to the other wall” sounds simple until the system has to decide which window, which wall, and whether the new location is allowed.
The ShedAi drafts describe a conversational layer connected to a 3D configurator. Its promise is useful: let people explain their intention naturally. Its implementation needs the same disciplined product rules as the buttons and forms.
Translate language into a proposal
Identify the intended object and the requested change. If either is ambiguous, resolve that ambiguity before mutating the design.
Represent the proposed action as structured data. Validate it against the current configuration and catalog. The model should not invent an option, a dimension limit, or a price because it sounds plausible.
Then show what changed and provide an accessible way to undo it. A visual highlight and a short explanation can keep the customer oriented.
Share one authoritative design
Direct manipulation and conversation should update the same configuration model. A door moved with a pointer must appear in the next conversational context.
Use stable object identities. Normalized wall coordinates can be useful for placement, but they do not independently enforce physical clearances. Evaluate the actual geometry and rules after a size change.
A camera-facing description such as “left” may need interpretation. A building’s documented orientation is a more reliable basis for saved specifications.
Keep business decisions outside improvised prose
Use authoritative calculations for price, availability, and permitted combinations. A narrative quote can explain a result, but the underlying amount should come from the approved pricing path.
Treat payment, order acceptance, and catalog publication as separate consequential actions. A conversational flow should not imply that a customer authorized a commitment merely by exploring a design.
Recover gracefully when the language service is unavailable. Ordinary controls should remain a practical route through the product.
Review imported catalogs
The archive also describes AI-assisted catalog extraction. An image or document can help propose a material, opening, or style, but inferred details need review.
Import only material the business is authorized to use. Record its source and distinguish measured dimensions from guesses. Validate units, duplicates, prices, and compatibility before publishing.
A picture of a window does not establish its rated properties. A successful scrape does not establish permission to reuse everything found.
Test the conversation in context
Try ambiguous references, unsupported changes, stale sessions, repeated requests, and mid-action failures. Check that a retry does not apply the same mutation twice.
Evaluate the result with real product tasks: can a customer discover a valid configuration, understand the changes, and finish without losing work?
The useful conversational system feels clear because its actions are dependable. Its language opens the door; the product model carries the decision through.
Keep a good idea close.
Follow ForgeWorks
