The core users Wegic targets do not necessarily have a design or engineering background. They might be coffee shop owners, independent photographers, freelancers, or people running small businesses. They may be able to clearly state what they want, but they generally lack the ability to judge the details of web design, such as whether a page's information hierarchy is reasonable, or if the brand temperament matches the business.

What they need is a complete service: clearly state what they want, and get a website that can be used directly and is of decent quality.

This means that the product must not only solve the "generation" problem, but also guarantee the final design quality. Below are my observations and judgments on several key stages of the generation pipeline's evolution, which I participated in within Wegic to "combat AI slop."

Section/Slot Stage: Transition Period's Splicing Scheme

Early code generation or Agent tools often showed mediocre output in the design dimension. Wegic initially adopted a splicing scheme of Sections and Slots—predefined modules internally, combined by AI based on requirements. At that time, Wegic did not yet have a full Agent form, but in that stage, this approach indeed provided relatively superior visual results from the user's perspective.

However, the internal perspective was: the AI was always rotating among a limited number of Sections. Many long-tail requirements could not be delivered, not because of model capability limitations, but because of physical bottlenecks in the scale and combination mechanism of the template library.

The sum of local quality does not equal overall quality. Although template splicing brought extremely low costs and fast response speeds, limited by the underlying splicing logic, the ceiling of this process was not high under professional design standards.

HiFi Prompt Stage: Stability Became the Problem Itself

After shifting to a real Agent form, we put a lot of effort into writing high-fidelity prompts, attempting to regulate the first-screen temperament, color strategy, card layout, and the restraint level of motion effects at the system level.

This set of methods is still effective today. The more detailed and precise the prompt, the more stable the output, and the higher the baseline completion rate of the design.

But this triggered another paradox: stability itself became the bottleneck. The more specific the prompt, the more converged the generated results. When spacing, rounded corners, and colors were locked down tightly, the AI's semantic understanding of specific projects was simply insufficient to support the differentiation of the final results. The demands of different industries and different brand temperaments ultimately led to visual convergence under strict text constraints.

Skill Stage One: Putting Prompts into Skills

When the Skill system was introduced internally, our first reaction was to encapsulate high-fidelity prompts into Skills. This is cleaner architecturally, and progressive disclosure also allows the system to extend into more style branches, partially alleviating the convergence problem.

But its underlying logic did not change: this is still a "locked-in style" mechanism, just with a different container for storage. Moving from one style corresponding to one package to one style corresponding to different variant packages of the same series, essentially what is still being maintained is a style template library.

Skill Stage Two: From Locking Style Back to Restricting Boundaries

The real turning point came after the multimodal models matured. Our early concept of "generating images first, then generating websites based on the images" finally became feasible.

Images can precisely convey sensory information that is difficult to describe in text: proportion, space, materials, light, and overall temperament. With images as visual anchors, the Agent's task during the building stage changed from "guessing design schemes out of thin air" to "high-fidelity restoration of visual intent." This process also handed design choices over to users upfront, allowing them to confirm the direction through low-cost "gacha" draws.

At this point, the role of Skills underwent a fundamental transformation. It was no longer responsible for "generating specific styles" but focused on two things: first, establishing negative boundaries to prohibit the AI from executing low-level designs (such as heavy shadows, inconsistent rounded corners, abrupt gradients); second, providing precise restoration capabilities after visual confirmation.

Skills stepped down from "defining styles" to "restricting boundaries and improving execution precision," which happens to be the area where this mechanism excels most.

Image-Driven: Currently the Best Path

The image-driven workflow builds a complete feedback loop: the AI generates multiple visual intent images based on requirements, the user confirms the scheme, and then the Agent performs a high-fidelity restoration based on the image. In this loop, users gain real decision-making nodes, the AI gains a visual anchor for comparison and regression, and the gap between delivery results and expectations is significantly narrowed. This is not a secret path, and it is indeed effective, but few products actually apply it. Perhaps, the transmission of visual intent should inherently be handled by the model's visual layer rather than the text layer.

But this is far from the end. Static images do not contain responsive rules, lack scroll rhythm design, and cannot define boundaries when content grows; even the aspect ratios of current image generation models often do not match real web scenarios. While it serves as a clever interactive strategy today that tangibly improves user satisfaction, fundamentally, it still cannot replace systematic, professional design decisions.

(This article only records my true perspective observing and participating in this evolutionary process within Wegic, and does not represent the company's official stance.)