Last CallLast Call Pool
Design Stage

What’s in your head - you can build. AI prototyping for UXers

September 18, 2025 11:40 AM
Long Lecture
🇬🇧 English
Wisła

About

Historically, building AI features has been an extremely expensive task. It required a team of engineers, data scientists, researchers and an enormous dataset. Now, that AI is on hype and we already have tools to experiment with and fine tune own models, UX designers still miss know-how on how to jump in and contribute to ai-related product work streams... and most importantly how to prototype desired experience.

This preso is intended to be a short intro to LLM finetuning and realtime onstage demo how to train own model and use it in simple html/css/js prototype or Protopie (tbc which one).

Talk by

Photo of Benjamin Shishko, Senior UX Designer at Google
Benjamin Shishko
Senior UX Designer at Google

Watch the full talk

Watch this WaysConf session, then continue with related talks or explore the current programme.

Explore the 2026 agendaMore WaysConf recordings
From the recording

Talk in brief

At WaysConf 2025, Benjamin Shishko, a Senior UX Designer at Google, demonstrated how user experience designers can leverage AI prototyping tools to build functional, highly personalized applications without deep coding skills. Shishko addressed three primary challenges in AI product design: generating reliable LLM outputs, prototyping unpredictable user experiences, and collaborating across cross-functional teams. Using Google AI Studio, he demonstrated a live build of a fitness application that utilizes structured JSON outputs, system instructions, and user-defined variables to generate tailored workout plans. He also highlighted Vertex AI as a robust platform for scaling prototypes into production-ready features through custom grounding, fine-tuning, and a mixture-of-agents framework. Shishko emphasized that UX designers must move beyond simple chat interfaces and rigid linear logic, instead using modular agents and continuous feedback loops to refine AI behavior and maintain brand consistency.

Key takeaways

  1. 01

    Controlling LLM Outputs with Structured Schemas

    Designers can use structured outputs to define a strict JSON schema, forcing the LLM to return data in a predictable format that fits directly into the application's user interface.

    Watch from 5:28
  2. 02

    Grounding and Fine-Tuning for Contextual Accuracy

    Rather than relying on public data, teams can deploy private grounding to search custom databases or fine-tune models with specific input-output pairs to prevent harmful or off-brand responses.

    Watch from 6:06
  3. 03

    No-Code Prototyping via AI Studio

    Google AI Studio allows designers to test prompts, compare model performance, and use natural language instructions to generate functional code prototypes without writing code manually.

    Watch from 19:13
  4. 04

    Decoupling Complexity with Modular Agents

    Instead of writing massive, single prompts, complex features should be broken down into a mixture of smaller, specialized agents where the output of one agent serves as the input for another.

    Watch from 29:30
  5. 05

    Designing for Non-Linear AI Logic

    Traditional step-by-step flows and rigid if-else logic fail in AI design because models dynamically select strategies based on user context, requiring designers to plan for flexible, non-linear interactions.

    Watch from 31:06
Read edited transcript highlights

These concise notes were edited from automatic captions and checked against the talk structure. They are not a verbatim transcript.

The Shift in Product Roadmaps

Shishko reflects on how the public release of ChatGPT in late 2022 dramatically altered product roadmaps. Features that were once confined to advanced research and development departments suddenly became top priorities for product teams. However, this rapid shift left product managers, designers, and engineers struggling to figure out how to collaborate effectively and leverage the technology. Shishko identifies the three main hurdles as ensuring reliable LLM outputs, prototyping highly unpredictable user experiences, and establishing cross-functional workflows to transition these features from experimental prototypes to production-ready quality.

Watch from 1:13

Technical Levers for Controlling LLMs

To build a reliable AI feature, such as a personalized fitness planner, designers must understand the technical levers available to manipulate LLM responses. Prompt variables allow user preferences to be injected dynamically into templates. Structured outputs enforce a strict schema, such as JSON, ensuring the model's response fits perfectly into the UI. Private grounding restricts the model's search space to a proprietary database, preventing generic internet hallucinations. Function calling enables real-time integrations like weather checks, while fine-tuning retrains the model on custom examples to enforce brand compliance and graceful fallbacks.

Watch from 4:13

Live Configuration in AI Studio

During the live demonstration, Shishko configures a chat-based workout generator in Google AI Studio. He explains key parameters like temperature, which controls the model's creativity, and thinking mode, which allocates a token budget for the model to plan its strategy before responding. He inputs a custom JSON schema to structure the workout days, exercises, and metadata. By using the platform's compare mode, he demonstrates how to run identical prompts across different models, such as Gemini 2.5 Flash and Flashlight, to evaluate response speed, schema compliance, and token consumption.

Watch from 10:31

No-Code App Generation and Iteration

Shishko showcases the 'Build' tab in AI Studio, demonstrating 'VIP coding' where natural language instructions generate a fully functional web application. He takes a basic fitness mockup and instructs the AI to replace dummy data with real backend logic connected to the Gemini API. The resulting prototype successfully processes user inputs—such as fitness goals, available days, and difficulty levels—to generate a structured workout plan. Shishko demonstrates the power of AI personalization by submitting a natural language request to remove walking exercises, which the model instantly processes to regenerate an updated, compliant plan.

Watch from 19:13

Transitioning to Vertex AI and Agent Frameworks

While AI Studio is excellent for initial prototyping, Shishko recommends Vertex AI for production-level development and team collaboration. Vertex AI allows teams to deploy robust retrieval-augmented generation (RAG) by indexing local files, reducing token costs and latency compared to stuffing databases into system instructions. It also supports fine-tuning workflows where UX designers and researchers can curate high-quality input-output pairs. For complex applications, Shishko advises decoupling massive prompts into a 'mixture of agents' framework, where specialized, smaller agents collaborate by passing outputs to one another.

Watch from 27:19

Designing Beyond the Chat Interface

In his closing remarks, Shishko challenges designers to think beyond standard chat interfaces and embrace structured data to create unique UIs. He warns against applying rigid, linear 'if-else' logic to AI experiences, as models are designed to dynamically select the best path based on user context. Designers should focus on modular agent architectures, continuous feedback loops to collect training data, and optimizing context costs. Finally, he stresses the importance of using shared workspaces so that cross-functional teams can collaborate on prompting, validation, and grounding in a unified environment.

Watch from 30:47
Explore WaysConf 2026

View the agenda

Get tickets