Appearance
AI as a Material: Modeling Behavior
The designer's goals haven't changed. The material has. Good design has always been about judgment, trust, control and outcomes. What's new is that the material can now hold judgment, behavior and autonomy itself. Interface design wrote scripts: every state specified, every path designed. An AI system doesn't run a script. It performs, more like an improv actor. So designers now write the rules, limits and intentions that the performance runs on. Designing an agent means designing its behavior, and that takes three things: your design intention, a clear description of the system, and a way to evaluate it.
What You'll Do
- Learn how a language model gets made, and where its behavior comes from.
- Compare the other kinds of models, and what each one learns to predict.
- Act out a letter-writing assistant with other people, and watch its behavior appear.
- Describe that behavior in a role card, then test it as a real agent in Agent Studio.
- Sort your design into the right place: system instructions, knowledge, or design notes.
Before You Start
- Workshop 1 helps. How It Actually Works explains what a language model does with your text. This workshop explains how that model was made. If you watched the 3Blue1Brown video there, ask yourself: what stuck with you, and what surprised you?
- For the acting exercise, you need two or three other people, some sticky notes or paper, and pens. Working alone? The Bodystorming an Agent guide gives you a way to do it with a chatbot.
- For testing, you need a computer with a web browser.
- No programming. Nothing here asks you to write code.
Two Ways Through
- For a quick run-through, follow the TLDR. It's one page of short steps, with links to the guides.
- For a deeper understanding, follow the guides below, in order. They explain each step, and end with questions to reflect on.
| TLDR steps | Guide | What it covers |
|---|---|---|
| 1–2 | How a Language Model Gets Made | Pretraining, fine-tuning and feedback, with real datasets, and how other kinds of models differ. |
| 3–4 | Bodystorming an Agent | Acting out a letter-writing assistant, then talking about what happened. |
| 3 | Bodystorming Kit | Printable goal cards, user notes, observer sheets and a blank role card. |
| 5 | Writing a Role Card | Describing the assistant's purpose, rules, boundaries, knowledge, inputs and outputs. |
| 6 | Testing a Role in Agent Studio | Turning your role card into an agent, and judging how it behaves. |
| 7 | Is This in the Right Place? | A checklist for sorting each sentence of your design. |
Then continue with Practice, before the next workshop.
Next
Workshop 3: AI as a Tool, where you build working prototypes by describing them, and give the tool your role card.
Examples
See prototypes made in this workshop on the Examples page.