← All episodes
Podcast
Ep #26: Paz Perez: Model Design, Context Architecture, and Designing for Agents
Tom and Jonny sit down with Paz Perez, a former architect and Google designer who coined the term model designer and now teaches context design. They dig into how designers shape data, context, and evals, explore context architecture and the future of design roles, and finish with quick-fire questions.
Key takeaways
- A first-ever Medium post gave a new discipline its name. After watching a Google content designer shape Bard with system prompts and rubrics, Paz wrote about the work and named it model design. The post went viral within a week or two, Google renamed a team after it, and job ads from OpenAI and others followed.
- Data, context, and evaluation are the new design materials. With no hard-coded step one, step two, step three, the output is shaped by training data, the context an agent can access (system prompts, knowledge bases, skills, tools), and constant evaluation. Designers bring the questions engineers skip: what is the user trying to achieve, and how do they feel along the way?
- Replace the linear funnel with an evaluation loop. Because outputs are non-deterministic, you evaluate continuously instead of once at the end. Paz teaches designers to think in loops of context, output, and review.
- Prototyping with AI is already context design. When you feed a design system to a coding tool, refine your prompts, and judge the result yourself, you are running the same loop with a human evaluator. Paz uses this to connect visually driven designers to the bigger AI product picture.
- Fast decision models make evaluations cheaper, with a transparency trade-off. A classification-style model outputs a decision rather than a paragraph, so it processes fewer tokens and is faster and cheaper, which suits pass or fail evals. The caveat Paz raises is that it behaves like a black box, so users see less of the reasoning.
- If context confuses a human, it will confuse an agent. With 200 skills and many connected databases, someone has to organise the context. Paz brought taxonomy, hierarchy, and labelling from information architecture into context architecture, and information architects in her class make agents more efficient and spend fewer tokens.
- Designers bring the moral compass to model choice. Models trained on the internet contain a tremendous amount of bad data. Designers can supply clean good and bad examples, red team for blind spots, and argue against a model when the use case could harm users.
- Specialist roles still matter, and complexity decides how much. Researchers supply the user understanding that gets distilled into system prompts, content designers organise written context, and service designers map where agents and humans interact. A small startup can have one person wear every hat, but a large complex product benefits from all of them.
- There is no learning path yet, and that is the opportunity. Paz has never taken a class on this and teaches what she has figured out, in a three-week course capped at 25 students. Her advice is to follow the right people, apply the ideas in your own products, and expect better classes in two years.
Chapters
- 0:00 — Introducing Paz Perez, who coined the term model designer
- 1:13 — From building public schools in New York to UX design
- 3:14 — Making the leap: a three-month bootcamp and why architecture prepared her
- 5:59 — Autodesk, conversational design, and natural language processing before LLMs
- 8:54 — The Google 20% project and the Model Designer article that went viral
- 13:15 — Teaching 900 Googlers about generative AI for UX teams
- 15:37 — Why the piece resonated with designers and content designers
- 17:35 — Would she write the article today, and designing for agents as personas
- 19:33 — Leaving Google, consulting on agentic workflows, and teaching
- 21:21 — Keeping a course current as the technology moves
- 22:50 — What a fast decision model unlocks for evals
- 25:30 — From designing pixels to designing context, loops, and evaluation
- 30:25 — Data for designers: golden examples, red teaming, and bias
- 33:13 — Research, content design, and service design in the model design era
- 37:31 — Context architecture and the return of information architecture
- 40:40 — How designers can prepare when there is no clear path
- 43:41 — Is this too technical? Why it adds to interface craft
- 46:46 — Is it like management? Collaborating with PMs and engineers on system prompts
- 49:39 — Working alongside her partner, an AI engineer
- 50:44 — Quick-fire questions: beaches, Moleskine, and favourite design tools
- 54:24 — Upcoming classes and where to follow Paz's work
Mentioned in the episode
The guest's work
- Model Designer article (Paz Perez, Medium)
- Paz Perez's Hatch talk on context, data, and evals
- Paz Perez's Maven course on designing context
- Context architecture article (Nielsen Norman Group)
- Information Architecture Conference keynote on context architecture
- GenAI for UX Teams (internal Google class)
Organisations and past roles
People
- Ovetta Sampson
Tools and products
About the pod
Mostly Working (formerly the Near Future Podcast) is a weekly conversation with design leaders about what they're building with AI. Send feedback to hello@mostlyworking.co, and find everything at mostlyworking.co.