Like with how I differentiate different components with different colors in the Paint program, as the components are being created by natural language instructions,
the program can label the parts separately for the user to better identify and communicate an edit. If that separation system is ambiguous at first, the user may select the areas to designate the parts with an label.

Current CAD programs rely on the memory of the user so these parts are not identified, you just see shapes. But the user will need to be able to differentiate components with a separate label of its own in order to better communicate to the A.I. Parametric (measured) numbers will show for every line that is selected or by default as well. This will better re-iterate the drag and shaping features of a CAD program replaced by human instructions, which in turn significantly reduce the amount of time that is required in design.

Next idea will be more difficult, but simulations, per type of simulation can follow this as well. However like with scripted instructions given to a worker who speaks through a phone to their customers in a business, there should be clear examples of communication. Though people can easily begin to work immediately without it, if the A.I has enough vocabularly and can adapt well into their instructions. Which is entirely possible. The script and specification is needed for more efficiency by the user in communication to the A.I.