AI Design: Two Techniques to Break Free from Mediocre Patterns

Precise methods are emerging to encourage a language model to produce something other than cloned pages. Between random prompts published by Sakana AI, highly directed prompts, and a three-step process, the method aims to surpass the predictable choices of LLMs without adding tools.

Introducing External Randomness with the "String Seed of Thought"

One approach involves injecting variety from outside the model. The technique called String Seed of Thought, published by Sakana AI, consists of having the AI generate a random string that then serves as inspiration for the design. This method is based on the observation that the model only predicts the most probable token and therefore cannot act in a truly random manner. By importing a source of randomness, each iteration can lead to different choices. This technique allows for the exploration of directions that deviate from average responses.

Formulating Ambitious Prompts Anchored in Strong References

An alternative is to provide the model with a clear, precise, and marked intention. Using references from personal experience, such as a video game, an interior design style, or an artistic work, can initiate the creative process. Among the suggestions are a pixel art design where each part recalls a game scene, an isometric 3D city illustrating different functions, or a highly asymmetrical composition mixing contrasting colors and fonts. It is acknowledged that finding original prompt ideas remains challenging, and simply asking the AI for ideas often leads to average proposals.

A Three-Step System for Generating Unique Prompts

A three-step process is proposed to improve the quality of prompts. First, ask the AI to list numerous ideas without going into detail, in order to stimulate imagination. Next, visualize the preferred options, note reactions, and ask the AI to refine these directions. Finally, iterate until satisfied, then request the AI to draft the final prompt. This framework structures the exploration while maintaining control over the creative direction.

Why LLMs Tend Toward Average Designs

Large language models learn to predict the continuation of a text sequence from vast corpuses of examples. Their outputs can be evaluated by humans, and these evaluations are reintegrated to guide them toward coherent and considered choices. This functioning explains good performance on many tasks but leads to very predictable design decisions. By proceeding token by token, the model favors what seems to please the majority, often resulting in repetitive and flat outcomes.

Field Observations: General Prompts vs. Cloned Results

With a generic prompt requesting a productivity app homepage, four instances of the same model produced nearly identical structures. According to feedback, Claude Opus 5 often generates a purple gradient background, a text area placed on the left, and a visual on the right. Even when formulating a request for a "totally unique" and "completely random" result, the observed diversity remains low: similar colors, recurring organizations, and close metaphors. These observations illustrate the interest in introducing a source of randomness or a more precise creative direction.

Proposed Process and Initiator Profile

The presented framework is based on three phases: Discover to explore avenues and formulate ambitious briefs, Define to establish an identity by pushing the AI out of its patterns through chaining models, and Deliver to polish a final output by focusing on key elements. This approach is rooted in a vision of design centered on emotion, rule-breaking, and unexpected choices. It is led by Anshu Chimala, who directed engineering and design teams at Apple for 12 years around the research and prototyping of future AI products, and who shares tutorials, analyses, and contacts on X, Substack, and LinkedIn.