AI in Creative Coding: From Completing Tasks to Crafting Narrative Experiences

Starting with Stone: Granting AI Creative Freedom

Past discussions around LLM-assisted development have primarily focused on engineering efficiency—completing functions, generating unit tests, and fixing compilation errors. In these scenarios, AI acts as an advanced auto-completion tool where humans define the problem and the model executes. However, I decided to conduct a more ambitious experiment: delegating the power of problem definition to the AI itself.

The goal was to create an interactive website that helps the general public understand how Michelangelo’s David was born from a block of stone. There were only three core constraints: the AI had to independently decide the narrative perspective, visual style, interaction method, and technical stack; it had to deliver a runnable, experiential product rather than just a proposal or mockup; and the audience should walk away with a new understanding of “sculpting” rather than just memorizing historical facts.

The core challenge of this task is that it is not an engineering problem with a standard answer. It is a creative problem requiring aesthetic judgment, narrative intuition, and user empathy. The model needed to strike a balance between accurately conveying facts and creating a unique experience.

Divergent Narrative Strategies: The Dark Room vs. The Exhibition Hall

The two models provided completely different answers, which is perhaps the most fascinating part of the experiment.

The core narrative of the first proposal was “subtraction.” It understood sculpting as “removing the unnecessary”—David was not added but “left over” within the stone. The entire website was designed as a dark room with a rough block of marble standing in the center. Users press and hold a button to simulate the act of striking with a chisel. As pieces chip away, David’s silhouette gradually emerges through repetitive subtraction. This approach intentionally leverages the physicality of interaction: every click is a sculpting action, and the press and release of the mouse simulate the striking of a chisel. This direct sensory correspondence turns the abstract concept of “sculpting” into a bodily memory.

Even more ingenious is the copywriting. The website features an “Undo” button, but when users try to click it, the prompt reads: “Marble has no undo key.” This design subtly conveys a philosophical viewpoint: sculpting is an irreversible decision-making process, and every strike permanently alters the artwork. This interaction pattern of “allowing the action but denying the result” is more powerful than any textual explanation.

The second proposal took the route of an exhibition catalog. It utilized a paper-colored background and divided the experience into a four-act structure, corresponding to different stages of sculpture creation. Users drag a slider to observe the stone block gradually revealing David’s form through layers of subtraction. This approach is closer to the narrative logic of a traditional museum—chapter-based, sequential, and emphasizing the systematic and complete nature of the information. The slider interaction allows users to move freely back and forth on the timeline, repeatedly observing the stone’s morphological changes at different stages.

The two proposals form an interesting contrast: one is an immersive experience emphasizing “presence” and “tactility,” while the other is an explanatory display emphasizing “structure” and “readability.” The former feels more like an artwork; the latter, a digital exhibition. Both utilized the same authentic 3D scan data of David from an open cultural artifact scanning project under the CC BY-SA 4.0 license, but handled the presentation in entirely different ways.

Technical Realities: Capability Boundaries and Collaboration Models

At the technical implementation level, this experiment also exposed some real-world conditions of current AI-assisted development.

First is the difference in model capabilities. The initial model performed outstandingly in reasoning depth and creative quality, but its speed and cost were less than ideal. During the website development process, generation was interrupted three times, ultimately forcing a switch to a lighter, faster model to continue. The transition after switching models was relatively smooth—the new LLM could read existing project files, understand the code structure and design intent of the completed parts, and take over the remaining work. This “mid-way substitution” collaboration model is extremely costly in traditional development but has become feasible in AI-assisted development, thanks to the model’s context window and its ability to understand project state, making the handover relatively seamless.

Second is the difference in execution efficiency. The finishing touches for the website were completed in six minutes, while the promotional video took about thirty minutes to produce. Such tasks often involve multiple stages—scriptwriting, visual generation, voiceover synthesis, and editing—each requiring the model to switch between different modalities, which significantly increases the time required. However, considering this was an end-to-end creative process starting from scratch—encompassing narrative design, 3D asset processing, frontend development, and video production—the overall efficiency still far exceeds traditional methods.

It is worth noting that the AI’s “autonomy” during the creative process was not uniform. At the decision-making level, the model demonstrated considerable judgment—choosing a dark room narrative over a traditional exhibition format, deciding to use “subtraction” as the core concept, and designing the “no undo” easter egg. These decisions show the model’s understanding of user psychology and narrative pacing. However, for visual details and interaction polishing, human intervention was still required. This suggests that the ideal paradigm for current AI-assisted creation is “humans define taste, AI executes the plan,” rather than a completely hands-off approach.

From Completing Tasks to Participating in Creation

Perhaps the most valuable finding of this experiment lies in redefining the hierarchy of AI’s participation in creation.

At the lowest level, AI is an execution tool—humans say “write a function,” and AI writes it. At the middle level, AI is a solution provider—humans say “build a website,” and AI provides the structural design. But in this experiment, the AI reached a higher level: facing an ambiguous, open-ended creative task, the AI not only independently completed the entire process from concept to delivery but also demonstrated narrative judgment at key decision points. It did not choose the easiest technical path; it chose the one most fitting for the thematic expression—the immersion of the dark room, the simulation of sculpting actions, and the philosophical metaphor of irreversibility. These are not the results of technical stacking but active choices based on an understanding of the theme.

Of course, this does not mean AI already possesses human aesthetic intuition. Its choices stem from learning and pattern matching against a massive corpus of existing works, rather than genuine “inspiration.” But when this pattern matching is precise enough, the output can trigger genuine emotional resonance. Many users experiencing the dark room version of the website reported that pressing and holding the button to chisel gave them a subtle sense of tension, and seeing the “Marble has no undo key” prompt made them smile knowingly while prompting deeper reflection.

Does this resonance come entirely from the AI’s design? Probably not entirely. It comes from the AI’s algorithmic judgment, the inherent appeal of the high-quality 3D scan assets, and the imagination the users project into the interactive space. Regardless, the AI plays the role of organizer and catalyst here—it weaves various assets and concepts into a coherent experience, and this weaving ability is a form of creation in itself.

Implications for Future Collaboration Models

Looking back at the entire process, there are several key takeaways for developers.

First, prompts for creative tasks require “delegation of authority.” Giving the model too many constraints stifles its judgment, while too few can lead to a loss of direction. A better approach is to clearly define the task’s goal and constraint boundaries while leaving the implementation path entirely to the model. The prompt for this experiment serves as a paradigm: it only specified “what we want the audience to understand” without dictating “what interaction method to use.”

Second, multi-model collaboration has practical value in real-world development. Different models make trade-offs across cost, speed, and creativity. Being able to switch models mid-project based on the nature of the task is a new paradigm for AI-assisted development. The key is that the project’s intermediate state—generated code and existing design decisions—must be understandable and resumable by different models, which demands higher context management capabilities.

Third, the availability of 3D assets and multimedia resources is a crucial prerequisite for AI creation. The success of this experiment heavily relied on the publicly available 3D scan data of the David statue. When models can access and utilize authentic, high-quality assets, their creative space expands significantly. This also suggests that in AI-driven creative projects, building a “supply chain” of assets is equally important.

Technology will not replace creation, but it is redefining the barrier to entry. In the past, building an interactive website like this required collaboration among designers, frontend engineers, 3D modelers, and content strategists. Today, a user with a clear goal and a sufficiently capable LLM can complete the entire process from concept to finished product in a matter of hours. The efficiency gain is only the surface; the deeper change is that everyone now has the opportunity to transform their own understanding and feelings into an experience others can engage with.

Just like that block of marble, what ultimately remains is precisely what was carefully chosen not to be chipped away.