How do you get people to trust an AI? I learned the rules designing Marty at Walmart, then used them to build and train Mandy, my own assistant, and a guest chatbot I designed and built for Vegas Suite Rentals.
AI assistants are easy to build and hard to trust. The technology can answer almost anything, which is exactly the problem: people don't know what it can do, whether it's right, or who is in charge when it isn't.
This is the story of how I learned to design for that, in three steps. At Walmart I helped design Marty, a generative AI partner agent in Seller Center, and came away with four rules. Then I used those rules to build Mandy, the assistant on this site, from interviews of my own work. Finally I used them for a chatbot for Vegas Suite Rentals that answers guests in the owners' voice. Each one taught me something the last one couldn't.
Marty at Walmart. A big team taught me what makes an assistant trustworthy.
Mandy, my own assistant. No team, so every decision and every wrong answer was mine.
Vegas Suite Rentals chat, a chatbot I designed and built for two suite owners. Would the rules hold up for their guests?
On Marty, I designed the experience and the states a conversation moves through, alongside product and engineering. On Mandy and the Vegas chatbot, I did all of it: the idea, the voice, the interface, the questions and answers, the links to my work, and the build. I connected both to Claude, Anthropic's AI model, and trained them by writing, testing, and rewriting their answers.
Where I learned the rules.
Marty sits in Seller Center. Sellers ask how to do something and get an answer that links to the right tool, with sources one click away and a thumbs up or down on every answer. I helped shape what Marty says and how it behaves, including researching what sellers ask and helping write the prompts that train it.
Working on Marty showed me what makes an assistant easy to trust, and it is mostly not the AI. It is what surrounds the answers: telling people it's AI before they start, and letting them check and push back on every answer.
Answers existed, but they were spread across Seller Center, and finding the right one took a seller who already knew where to look.
Guides, Seller Academy, and support each held part of the answer, so no one place had all of it.
Sellers had to already know where to go in Seller Center to find the tool they needed.
Simple questions often went to support instead of being answered where the seller was working.
A chatbot is only as good as what it is taught to say. Working closely with engineering, I did a lot of research into what sellers ask and where the answers live, and I helped write the prompts that train the chatbots.
What I researched
I looked at the questions sellers were asking and the support tickets they turned into, to help the team see which problems came up most and how sellers phrased them in their own words.
I went through Seller Center, Seller Academy, and the guides to help map where each answer lived, where the content overlapped, and where it was missing, since Marty could only be as complete as the sources behind it.
I talked with the people closest to the problem, including sellers, support, and subject-matter experts, to help the team understand what a good answer looks like and what a wrong one costs.
I looked at how other chatbots handle answers, sources, and mistakes, and what current AI best practices say about being clear on what the bot can and cannot do, and brought that back to the team.
What the prompts covered
I helped set how Marty sounds. Plain, helpful, and in the language sellers use, not the language of internal documentation.
I helped shape the answers themselves: short responses, numbered steps when a seller asks to be walked through something, and links to the Seller Center tool that finishes the job.
I helped define what Marty should decline, when it should send a seller to support instead of guessing, and how it handles uncertainty, so a wrong answer never looked like a confident one.
Alongside engineering, I helped run sample questions against the prompts and rewrite them based on what came back, repeating until the answers held up across the kinds of questions sellers actually ask.
Four steps take a seller from a question to an answer they can check.
Sellers type it the way they would ask a person.
A “Thinking…” state shows Marty is working on it.
With links to the Seller Center pages they need.
Sources are one click away, plus a thumbs up or down.
Six decisions shaped how Marty behaves, from what sellers see before their first question to what they can do with the last answer.
Terms of Use before first use, and a short reminder under every chat box.
Sample questions and topic buttons on the welcome screen.
A “Thinking…” state shows sellers it’s still working.
Answers link to pages like Media Library and Add Items.
Ask “walk me through the steps” and get a numbered list.
Thumbs up or down and sources on every answer.
Before a seller asks anything, Marty asks them to read and accept the Terms of Use. After that, a short reminder under every chat box says that AI can make mistakes and not to share sensitive information.
Here is one question from start to finish. Each step tells the seller what Marty is doing and what they can do next. It plays on its own, and you can step through it yourself.
I handed off the flows for each state, so engineering knew what to build and how each answer should behave.
Marty is designed to give sellers and suppliers one place in Seller Center to ask how to do something, with answers that link to the right tool, sources one click away, and a thumbs up or down on every answer. The goal was for simple how-to questions to be answered where the seller is already working, instead of becoming support tickets.
I contributed to the design alongside product and engineering, from the Terms of Use through the loading, error, and step-by-step states. I moved on before launch, so I can speak to the design, not the results.
Four rules came out of Marty. They are the rules I carried into everything after it.
People should know they are talking to AI, and that it can be wrong, before they rely on an answer.
A wrong answer that sounds sure is worse than no answer. Sources and links let people check.
People can rate an answer, push back, or reach a human. The assistant never has the last word.
An assistant that knows what it can answer, and where to send everything else, is easier to trust than one that tries everything.
Those four rules are what I took with me. The next question was whether I could apply them on my own, from start to finish.
Where I used the rules on myself.
Mandy lives on this site. Visitors can ask her about my work, background, and process, and she answers in short, linked bullets, at any hour.
I wanted a real, working example of how I design AI, not just a description of it. Marty had a large team and a long review process. This time there was no team and no review. Every decision, and every wrong answer, was mine.
This is the real assistant. Ask her something, or switch to the hiring agent: tell it your company and the role, and it researches, drafts three reasons you need the role, reads my recommendations, and prepares an offer. Then it stops and waits for you to approve or decline. It is scripted on purpose, so it never does anything real.
Nothing you type is saved.
The tags show which of Marty’s four rules each decision came from, and which are new from using her.
She introduces herself as my AI, and says I designed, built, and trained her, before anyone asks.
Starter questions and topic buttons on the welcome screen, so the first question is a tap.
A lead-in and a few bullets are quick to scan. Walls of text are not.
Answers point to the page and section on this site, so a visitor can check the work themselves.
Closing the chat clears it. Nothing a visitor types is kept.
"Contact Amanda" stays at the bottom of the chat for anything she should not answer.
Design only gets you the shell. What Mandy says had to come from me, so I made her by interviewing myself, then turning those answers into an assistant I trust to speak for me.
I recorded interviews of myself answering the questions a recruiter or hiring manager would ask, in my own words.
I imported the transcripts and the questions into the agent, so her answers start from how I really talk about my work.
I connected the agent to Claude, with a small server file that keeps the AI key off the page.
I asked her questions, read what came back, and changed the answers wherever they missed, again and again.
Training Mandy felt less like coding and more like coaching a new hire who has read my whole career but has never met me. The interviews gave her my words. Everything after that was editing: deciding what good looks like, then saying it clearly enough that she does it every time. It took a lot of hands-on work to make it feel right.
She answers from what I told her in the interviews. She is told never to invent employers, dates, metrics, tools, or opinions, because a made-up detail about my career would do real harm.
I linked images and projects to questions, so an answer about a project can point to the exact page and show what I made, not only describe it.
Every answer opens with a short lead-in, then two to five short bullets, one idea each. She is warm and professional, and she sounds like one assistant every time.
She stays on my work, background, process, and how to reach me, and politely declines the rest. She does not share confidential details from past employers, and she does not discuss salary or whether I would accept an offer.
The early answers were too long and repeated themselves. Fixing that was most of the job. I would ask a question, read the answer like a critique, decide what was wrong, and change it: never write a paragraph, never repeat yourself, never ask more than one question. Each fix is small, but together they are the product.
It is the same skill as design critique. You look at what came back, name what is off, and say what you want instead. The difference is that the assistant does the same thing every time, so a small, clear fix improves every future answer.
An assistant that speaks for me on a public site needs limits. The AI key lives in a separate server file, never in the page. Mandy only answers requests that come from this site, and there is a daily cap on messages so cost can't run away. She runs on Claude Haiku, a small, fast model, which suits short, quick answers.
A chatbot I designed and built for two suite owners.
Guests of Vegas Suite Rentals ask the same questions again and again: parking, check-in, pets, fees. I built a chat on the guest site that answers them at any hour, in the owners' own voice.
It introduces itself as "Pat & Bernie, Superhosts on Airbnb," and is styled to match the dark and gold look of the site. Read the Vegas Suite Rentals case study.
What guests can ask it
The four rules I learned on Marty held up every time, whether the assistant was for a large company, for me, or for two suite owners.
Marty shows Terms of Use and an AI reminder before the first question. Mandy introduces herself as AI in her first line. The Vegas chatbot introduces itself up front and invites the first question.
Marty puts sources one click away. Mandy links to the exact spot, and the hiring agent shows every step it takes.
Marty lets sellers rate answers and push back. Mandy asks you to approve the offer and always offers a way to contact me.
Marty has guardrails for what to decline. Mandy only talks about my work. The Vegas chatbot sticks to questions about a stay.
Looking back, I learned that trust is a design problem. Saying what the assistant is, showing where answers come from, and keeping a person in charge did more for trust than making the assistant sound confident.
I learned that training an agent starts with real material and ends with editing. Interviewing myself gave Mandy my words, and the work after that was reading her answers closely and rewriting the ones that missed.
I learned that agents need a visible plan and a human decision. The hiring agent works because you can see each step and nothing happens until you say so. That is the same idea behind Marty's sources and thumbs up or down.
Assistants and agents are the kind of problem I look for. A new technology that has to feel clear and honest to the people using it.
I like designing for the moment someone needs one clear next step, not another place to search.
AI is new, but sellers want a clear answer, so I turn it into states and answers they already understand.
I like work where honesty shows up in the interface, like what Marty is, what it's doing, and where its answers come from.
I designed the experience for Walmart’s AI assistant, then used what I learned to design, build, and train two more of my own. I know how to make an assistant clear, honest, and safe in front of real people, and I use AI in my own design work every day.
If your team is bringing an assistant or an agent to the people who use your product, I'd like to help make it trustworthy.