I use ChatGPT extensively to develop scientific software, including browser-based interfaces as well as native Python/C++/CUDA applications for GPU-accelerated image processing, 3D visualization, scientific analysis, and data conversion.
My workflow has exposed two areas where ChatGPT could dramatically improve software development for scientists, engineers, researchers, and other domain experts who are not full-time software developers.
1. Direct click-to-edit interaction with generated app/site previews
ChatGPT can already generate an application and show an interactive preview. I would like to be able to click directly on an element of that running application, such as a button, panel, plot, histogram, 3D viewer, text field, or control, and then tell ChatGPT:
“Make this larger.”
“Move this outside the graph.”
“This control isn’t working.”
“Change only this section.”
ChatGPT would know exactly which UI component I selected and modify the corresponding code.
This would make software iteration much closer to the way image editing works: point at the thing, describe the change, and let the model determine the implementation.
2. Install once, then let ChatGPT update native applications in place
Browser-only applications already have a relatively smooth iteration cycle. The larger problem occurs with applications that require native components such as Python packages, C++, CUDA, GPU libraries, local file-system access, or other dependencies.
My current workflow is roughly:
ChatGPT → generate version → download → install → run → test → screenshot/report problems → ChatGPT modifies it → download another version → reinstall → test again.
During R&D this can happen dozens of times.
I would like ChatGPT Work/Codex/GitHub/local development tooling to support an install-once development application.
After the initial installation, ChatGPT should be able to modify the project, rebuild only what changed, update the existing development installation, restart/reload the necessary components, and preserve configuration and dependencies whenever possible.
Ideally the user experience becomes:
Describe change → ChatGPT edits code → build/test → update existing app → user tests → repeat.
Rather than:
Describe change → rebuild package → download → uninstall/reinstall → configure → test → repeat.
For managed corporate/scientific computers, this would be particularly useful if the workflow supported user-space installations and updates that do not require administrator privileges whenever technically possible.
Why I think this matters
I am a scientist rather than a professional software engineer. I understand the mathematics, physics, scientific workflow, experimental requirements, and what the end users actually need. I increasingly use ChatGPT as the implementation layer between that domain knowledge and working software.
I don’t necessarily want to become an expert in installers, build systems, dependency management, CUDA packaging, CI/CD, or deployment infrastructure just to translate a scientific idea into useful software.
I believe there is a potentially important class of ChatGPT users like this: domain experts who can design and validate sophisticated software but do not want software engineering itself to be their primary profession.
Combining direct manipulation of app previews with an install-once/update-continuously native development workflow could turn ChatGPT into something closer to a complete iterative engineering environment for these users.
I have attached a more detailed product proposal describing the workflow, architecture concept, use cases, and implementation ideas.