Many AI design experiences stop inside a chat box.
You describe a spatial problem, and the AI gives advice. You add a few images, and it analyzes them. You ask for another style, and it writes a new direction. In the short term, this is useful. But once a project runs for days or weeks, the chat box starts to show its limits.
Advice becomes scattered. Versions blur. Source material loses context. A decision that made sense yesterday may be impossible to explain tomorrow.
A real design project needs more than advice. It needs a durable project surface.
Why chat advice is not enough
Chat is a good thinking interface, but it is not a stable project container.
It has three natural weaknesses.
First, chat history is a weak source of truth. One decision may hide inside ten messages, one dimension may come from an image, and one rule may be a temporary instruction you typed in a hurry. The next time AI makes a change, it can easily miss one of those facts.
Second, chat history is weak at versioning. You can tell AI to "change it according to what we just discussed," but the project needs to know what changed, why it changed, whether it affects other spaces, and whether it can be rolled back.
Third, chat history is weak at holding material. Drawings, references, components, materials, screenshots, feedback, and rules all need organization. If they are only dropped into a chat window, it becomes difficult to review provenance and validity later.
If an AI design tool only optimizes chat, it becomes a smart but forgetful consultant.
The project workspace is the ground of the workbench
A more reliable approach is to put each design task inside its own project workspace.
That workspace is not just a folder. It is the shared operating surface between the designer and AI.
It can hold:
- the current design model;
- project rules;
- source drawings and material;
- components and material lists;
- screenshots, renderings, and review records;
- version differences;
- temporary memory valid only for this project;
- next actions.
With this structure, AI does not have to infer project state from chat history every time. It can read explicit files, modify explicit files, produce explicit artifacts, and write important changes back to the project.
For designers, this is closer to real work. A design project is not completed by one sentence. It is made of material, rules, models, drawings, communication, and versions.
A durable project needs structured truth
A project workspace is still not enough. It needs a clear structured truth layer.
In a direction such as SketchUp Agent Harness, that layer can be design_model.json. It records spaces, dimensions, components, rules, assumptions, and execution state.
The point is not that designers should hand-write JSON. The point is that AI, the designer, and the software need a shared fact layer.
If the design only lives inside a SketchUp scene, AI has a hard time knowing which parts came from source evidence and which parts are temporary operations. If the design only lives in screenshots, it is hard to repair. If the design only lives in chat, it is even harder to compare versions.
A structured truth layer lets the system answer:
- what spaces and components exist now;
- where the dimensions came from;
- which assumptions are still unconfirmed;
- which feedback has been accepted;
- which rules affected this change.
That is the difference between a durable project and one-off advice.
AI should turn conversation into project state
Chat still matters. Designers should be able to say naturally: "This entrance is too narrow. Move the cabinet beside it a little."
The key is that AI should not only reply, "Sure, I will adjust it." It should turn that sentence into project state:
- identify the affected space and component;
- check whether project rules allow the change;
- create a modification plan;
- update the structured model;
- execute into SketchUp or prepare execution;
- leave a change record;
- generate a screenshot for confirmation when needed.
Then conversation does not evaporate.
Each useful exchange should move the project state forward.
A good workbench lets projects pause and resume
Real projects are often interrupted. You work on one version today, switch to another client tomorrow, and return three days later.
If an AI design tool depends only on chat memory, resuming the project becomes painful. You have to explain the background again, upload material again, and remind the tool of the rules again.
A workbench-style tool should let the project pause and resume. When the designer opens the project again, AI can see the current model, rules, material, snapshots, accepted feedback, and unfinished tasks.
This is not about giving AI infinite memory. It is about keeping memory in the place where it belongs: the project.
Over time, the real value of an AI design tool is not making every chat answer prettier. It is turning chat, material, model state, and feedback into a design project that can continue.
Originally published on my personal site:
https://marlinbian-site.pages.dev/writing/from-chat-advice-to-durable-design-project/
More links: GitHub · YouTube · LinkedIn · Bluesky · Mastodon · Discord
Top comments (0)