Harbor Bai   
   
          Selected Works             About                                 CV


Prompt-Sculpting Controller:




  • Role: 
    End-to-end product designer/engineer (solo)
    Full stack web developer
  • When:
    Oct–Dec 2025
  • Focus: 
    Tangible UI/UX design
    Computational creativity 
    AI human interaction

  • Tools: 
    Rhino/Fusion 360, Arduino (firmware),
    Node.Js
    Cursor 
    vite


    • Overview:
    While LLMs are powerful creative partners, typing remains a limited input method—discrete, linear, and symbolic—unlike human thought, which is high-dimensional and non-linear. This creates a gap between how we think and how we must communicate with AI.

    This project explores how physical interaction and spatial visualization can reduce this gap. By treating prompt creation as a performance rather than a writing task, the Prompt-Sculpting Controller aims to expand expressiveness and lower the cognitive burden of language manipulation.





    Background & Motivation:
     LLMs and Transformers:

    Transformers interpret language through attention patterns, semantic embeddings, and token relationships—not simply through the linear sequence of words. However, typing forces users to encode their complex intentions inside this linear stream. (Soni, 2024)



    BERT:



    BERT (Bidirectional Encoder Representations from Transformers) is a transformer-based language model that encodes language bidirectionally, allowing O₂ to interpret user prompts as rich semantic structures rather than linear text, bridging the gap between human intention and machine understanding.



    Grammar as a Constraint:

    Traditional sentence structure (e.g., subject → verb → object) imposes a sequence not always aligned with human conceptualization (we don’t think linearly). Reed–Kellogg diagrams provide a structural view that externalizes grammatical relationships. (“Reed-Kellogg Diagrammer Help,” n.d.)




    Design Concepts:

    The Prompt-Sculpting Controller reimagines prompt creation as an embodied, musical interaction.
    The system splits language into two modifiable dimensions:
    • Semantics (meaning variations)
    • Syntax (order and structure)
      Users sculpt a sentence’s latent properties using continuous rotations, similar to modulation in musical performance.

    The interface shifts prompting from “typing text” to “exploring a field of meaning.”




    CAD:



    Full stack web app development:

    The UI integrates Reed–Kellogg diagram concepts into a radial layout, translating syntactic hierarchy into interactive geometry.



    Assembly:



    Contributions:
    This work contributes:
    1. A new tangible interface for LLM prompt design, replacing linear typing with gestural modulation.
    2. A spatial representation of sentence structure inspired by Reed–Kellogg diagrams.
    3. A bidirectional physical–digital pipeline connecting embodied interaction with transformer-based language generation.
    4. A new interaction paradigm: “performative prompting,” where users sculpt meaning continuously.





    Workflow demonstration：




    Conclusion：
    The Prompt-Sculpting Controller demonstrates an alternative to typing for interacting with language models. By externalizing grammar, spatializing meaning, and adding gestural expressiveness, it transforms prompt engineering into a tactile, creative, and intuitive process.






    Harbor Bai © 2025. All Rights Reserved