Will multimodal AI reshape the software development process?

Looking at the speed of evolution of these multimodal AIs in Google Gemini 2.5 Pro, I can't help but wonder: is software development really going to be completely rewritten? The traditional linear process of "write code - debug - release" is being torn apart by AI. Imagine, the product manager draws a sketch, the designer throws a PSD file, the marketing department gives a product video, AI can directly spit out the code prototype can be run - this "what you see is the code" ability, directly to the product development cycle cut a big half. This "what you see is code" capability directly cuts the product development cycle in half.

But don't think it's just a simple code generation tool. Anyone who has tested Gemini 2.5 Pro recently has the same feeling that the most frightening thing about it is its "contextual understanding". For example, if you let it change the style of a button, it will actually take the initiative to check the consistency of the entire UI style; when dealing with the error report, it will not only give solutions, but also analyze the chain reaction that may be triggered. This kind of systematic thinking close to human engineers is what traditional low-code platforms can never do.

The developer's new dilemma: from artisan to curator?

There's an interesting paradox: while GitHub Copilot made developers 55% more efficient, Stack Overflow's traffic dropped 15%, and while AI is starting to understand the fuzzy need to "take the interaction in the video on the left and put it into the table on the right," programmers' knowledge systems are being reconfigured. The programmer's knowledge system is being reconstructed. A while ago, a full-stack engineer friend complained that the time he spends on adjusting prompts has exceeded the time spent on writing code. This may indicate that the future developers are more like "demand curators", and the core skill becomes translating business logic into multimodal prompts that AI can understand.

However, do not be too optimistic, the actual test found that this type of tool in the design of complex systems will still be overturned. Last week, I let Gemini 2.5 Pro generate an e-commerce app with payment function, the front-end interface is amazing, but once I checked the ordering process, I found that the inventory validation logic is flawed. This shows that AI is currently better at "module-level innovation", and enterprise-level development of the most important system reliability, or the need for human engineers to check.

Chain reaction of the development process

The changes brought by multimodal AI are chained. When the UI design can be directly converted into code, designers will have to consider the feasibility of development; when the product demo video can be broken down into function points by AI, the PRD document may have to be rewritten. It has been seen that some startups have merged the product, design and development teams into a "digital product creation group", where all the people work around AI tools in real time. This mode of work, put three years ago simply unimaginable.

There's another trend that can't be ignored: AI is eating unit testing for lunch. With the ability to understand natural language descriptions such as "this button click should send a request instead of refreshing the page", the efficiency of test case generation has increased by more than 8 times. But this also brings a new problem - how to ensure the validity of the test when both the test code and the implementation code come from the same AI? The solution of a major first-tier manufacturers is to require AI to use different models to generate realization and test code, this "left and right fighting each other" mode is quite inspiring.

At the end of the day, multimodal AI won't put programmers out of business, but it certainly puts programmers who don't know how to use AI out of business. Just like Photoshop didn't eliminate designers, it just eliminated designers who can't use PS. The future of software development may be a combination model where humans are responsible for drawing creative boundaries and AI is responsible for exploring efficiently within them. You get everyone on your team to sign up for a Gemini 2.5 Pro account right now and try it, and you're guaranteed to hear about all sorts of amazing workflow innovations at next week's stand-up meeting.