How I AI: Priya Badger's Workflow for AI Product Management and Prototyping at Yelp

Priya Badger, a product manager at Yelp, starts conversational product work with "golden conversations" before anyone writes a PRD or builds a polished interface. The sample dialogue becomes the first prototype. From there, she uses AI tools to pressure test the interaction, generate system instructions, and explore what the surrounding product experience should look like.

In this episode of How I AI, Priya walks through a workflow that starts with the end-user conversation and works backward into requirements, prompting, and interface design. Instead of treating conversation design and product design as separate activities, she uses AI to move between them quickly and test ideas while they are still cheap to change.

Priya’s team at Yelp is building an assistant that helps people find service professionals such as plumbers, electricians, and handymen. A user describes a problem, the assistant asks follow-up questions, and Yelp helps connect them with pros. Adding photo uploads created a harder product question: once a user uploads an image, how should the assistant change its questions, recommendations, and flow? Priya’s process starts with example conversations, turns them into an interactive prototype, and then tests whether the behavior still works across very different service scenarios.

Designing AI products from "golden conversations"

Priya starts by writing example conversations that represent the intended user experience. Yelp internally calls these "golden conversations." Instead of debating requirements abstractly, the team can react to a realistic interaction and decide what the assistant should notice, what it should ask next, and how concise or opinionated it should be.

Generating the first conversation draft in Claude

Priya opens a fresh Claude chat and asks it to role-play a conversation between a Yelp user and the assistant. The scenario centers on a customer uploading a photo for a home-service request. She notes that the same approach works in ChatGPT or other models, but Claude is useful later because of its built-in prototyping features.

Her prompt is detailed and structured rather than conversational. She specifies the roles, the formatting style, and the kinds of behaviors the assistant should demonstrate so the output is easier to evaluate and iterate on.

write a complete sample conversation between the consumer and AI assistant, where we want consumers to be able to upload their photo. Add some scenario requirements, like we want the assistant to analyze the photo, maybe provide some suggested replies, and continue that back and forth until they have enough info to submit quotes. Use assistant colon, user colon for labels. Write it as one continuous conversation.

That structure matters because it gives Priya something concrete to critique. Instead of discussing vague feature ideas, she can react to an actual interaction and refine the experience directly. I compare the method to writing developer documentation before building a developer tool: work as close as possible to the thing the user will ultimately experience, then back into the implementation details later.

Testing the conversation against real images

After generating the initial dialogue, Priya uploads a real image of a cracked porch to see how Claude handles the interaction in practice. She pays attention to what the model notices in the image, which assumptions it makes, and whether the conversation feels useful or awkward. Claude’s visible reasoning process also helps her debug prompts and understand where instructions are too vague. She treats that reasoning view as a troubleshooting aid, not as access to hidden internal reasoning.

Claude correctly identified the visible crack in the concrete steps and asked follow-up questions about urgency and whether the user wanted repair or replacement. Priya was not evaluating the structural advice itself. She was evaluating the product behavior: did the assistant ask sensible next questions, did it understand the image, and did the conversation move naturally toward getting quotes from professionals?

One successful example is not enough for a product that spans dozens of service categories, so Priya expands the test set with several very different images:

Now create more examples based on these images. Label each conversation with a title and a number at the top.

She uploads photos of an appliance displaying an error code, a hornet or wasp nest, and a bathroom renovation in progress. The point is not just image recognition accuracy. She wants to see whether the assistant behaves consistently across wildly different homeowner problems and whether the conversation style still works when the context changes.

At this stage, Priya reviews the outputs qualitatively rather than scoring them with a formal evaluation framework. She looks for recurring patterns: responses that are too long, questions that feel repetitive, situations where the model misses obvious visual details, or moments where the assistant should sound more decisive. Those observations later inform a more structured evaluation rubric.

Refining conversations and turning them into prototypes

Priya then polishes the conversations based on what she’s learned. For example, if she wants the AI to have a stronger opinion or to avoid asking about budget (since a user might not know), she can just tell Claude to rewrite these conversations based on this feedback.

The examples also help Priya identify where the assistant should express stronger guidance. In the wasp-nest scenario, she pushed the model to be more direct, and the prototype recommended professional pest control instead of attempting removal. It is a good example of where product judgment matters. The assistant may need to sound confident in safety-sensitive situations, but those behaviors still require expert review and careful testing before becoming part of a real product.

Once the conversations feel solid, Priya turns them into an interactive prototype using Claude Artifacts. At the time of recording, Artifacts could generate a working chat interface backed by Claude itself without requiring separate API setup or custom infrastructure.

Her prompt asks Claude to create a chat application, generate the underlying system instructions from the sample conversations, support photo uploads, and visually resemble the existing Yelp Assistant experience.

create an assistant app as an artifact, have a chat interface where the AI responds using the LLM that powers Claude. And then also create system prompt that is based on these example conversations. And then analyze these, uploaded photos, and include a camera icon in the input.

Priya also uploads screenshots of the existing Yelp Assistant front-end to help guide the visual style of the artifact.

One useful side effect of this workflow is transparency. The artifact exposes the generated system instructions, so Priya can inspect how Claude translated the example conversations into behavioral rules. She uses that output both as a prototype and as a learning tool for understanding prompt and system-design patterns.

When Priya tested the prototype with an oven photo showing an F2 error and a locked door, the assistant recognized the visible issue and offered troubleshooting guidance. The interaction helped validate the conversation flow, but she still treats appliance and safety guidance carefully. A polished interaction does not remove the need for manufacturer instructions, escalation paths, or qualified professionals.

The interactive prototype exposes problems that static documents hide: mobile pacing, response length, waiting states, and awkward follow-ups. The golden-conversation workflow shows how Priya moves from example dialogue to system instructions and a testable chat artifact before implementation.

Exploring interface directions with Magic Patterns

After defining the conversational behavior, Priya switches to Magic Patterns to explore the surrounding interface and user journey. She recreates the existing Yelp Assistant UI first so new ideas appear inside a familiar product context rather than as disconnected mockups.

Adding a photo-upload entry point

Priya starts by recreating the Yelp Assistant UI in Magic Patterns, which gives her team a familiar canvas to work on. Her first task is to think about how a user would even start the photo upload process.

Priya prompts Magic Patterns to add a more visible entry point for photo uploads and describe how the flow should work inside the assistant experience.

add a prompt suggestion at the top for start with a photo, which allows the user to upload a photo.

The generated result gives her a concrete design direction to react to: a camera icon, instructional copy, and an updated layout around the chat interface. Priya emphasizes that this is not production-ready UI. It is a fast exploration surface that helps the team discuss possibilities earlier in the process.

Using Inspiration Mode to generate divergent ideas

Priya uses Magic Patterns’ Inspiration Mode when she wants broader exploration instead of small tweaks. She specifically prompts it to create differentiated concepts so the output does not collapse into four near-identical layouts.

Her prompting strategy matters here too. She explicitly asks the tool to think expansively and explain the reasoning behind each variation so the team can compare product directions rather than cosmetic differences.

Give me some options on how to start with the photo flow could work to make it feel more guided for the user. Think expansively and make each option differentiated. Explain in your response what each option is.

Magic Patterns then creates several different UI options and shows them side-by-side. Priya can click through them to see right away how different design ideas would look for the start with a photo flow.

The generated options included ideas such as guided category selection and a more real-time image-detection flow. One variation failed during the demo because of a React compatibility error, which Priya and I both treated as a normal part of AI prototyping. Even with the failure, the exercise surfaced several distinct product directions worth discussing and refining.

Priya describes these tools as collaborative accelerators rather than replacements for design or engineering. The prototypes help product managers communicate ideas visually and interactively, but the team still has to decide which concepts match the user problem, technical constraints, accessibility requirements, and broader product strategy.

Using the same workflow on personal projects

Creating a community newsletter from Slack discussions

For a recurring internal AI newsletter, Priya saves reusable instructions inside a Claude Project so she does not need to rewrite prompts each time.

I'm a community manager writing a weekly newsletter. Use these Slack conversations and format them just like the Community Wisdom Newsletter.

She exports Slack discussions, replaces or removes names for privacy, and asks Claude to summarize the recurring themes into a newsletter draft. The workflow dramatically reduces the effort required to synthesize large community conversations, but Priya also notes the operational side of the process: teams should only export approved data, remove sensitive information, and follow internal retention and access policies.

Building a parenting assistant with Claude Artifacts

Using the same Claude Artifacts workflow from her Yelp project, Priya built a Parent Pal. This is her personal AI assistant that helps her and her husband with common parenting challenges, like what to do about a 2-year-old throwing toys down the stairs.

Using the same Claude Artifacts workflow she demonstrated for Yelp, Priya built a small "Parent Pal" assistant for handling common parenting situations. In one example, the assistant asks follow-up questions about a toddler throwing toys down the stairs before suggesting possible responses. The project is intentionally lightweight and personal, not a replacement for professional medical or developmental advice.

Prototyping a Settlers of Catan timer

Priya's siblings love playing the board game Settlers of Catan, but games can run long. She's building a Settlers of Catan timer in Lovable to help. The project is growing based on feature requests from her siblings, who now want tracking wins, a leaderboard, and handicaps.

Priya is also building a Settlers of Catan timer in Lovable because games with her siblings tend to run long. The project started as a simple timer but expanded after family members requested leaderboards, win tracking, and handicaps. For aspiring AI product managers, these side projects are useful because they force you to define behavior, handle edge cases, and iterate on product ideas even if your day job is not yet AI-focused.

Start with the conversation and force the product to prove itself

Priya’s workflow flips a common product-management instinct. Instead of beginning with requirements documents or polished wireframes, she starts with the user interaction itself. The conversation becomes the source material for system instructions, prototypes, and interface exploration. Claude Artifacts helps her test whether the interaction actually works, while Magic Patterns helps the team explore how users discover and move through the feature.

What makes the workflow useful is not that AI generates everything automatically. It shortens the distance between an idea and a realistic interaction. Designers, engineers, and PMs can react to the same artifact much earlier, while changes are still inexpensive and the product direction is still fluid.

The most transferable part of Priya’s process is the discipline of writing concrete example conversations before building. It works especially well for conversational products, onboarding flows, customer support experiences, and multimodal assistants where tone, pacing, and follow-up questions matter as much as functionality. The prototypes are fast and surprisingly persuasive, but they still require human judgment around safety, escalation paths, accessibility, and what the product should or should not confidently say.