Case study

Designing AI agents for HBO Max

I designed and built the first generation of agents that write personalized content descriptions based on a user’s watch history.
In early testing, 54% of test users said the descriptions helped them decide whether to watch.

RoleAI Product Design Intern
Timeline8 weeks
Team1 UX writer, 1 senior
designer, 1 UX researcher
StatusHanded off
The HBO Max TV home screen featuring House of the Dragon, with the short description under the logo circled.
Handwritten content descriptions will be replaced by AI-generated ones.

01 — Problem

Viewers browsed, but didn’t press play.

Summer 2025 Findings

  • 2 minusers spent looking for a title
  • 5 of 7leave without playback

HBO Max had plenty to watch, but nothing on the screen convinced viewers to press play. I saw the descriptions as a place of opportunity. I wanted them to grab viewers’ attention and give them a reason to watch.

02 — Idea

The ask was to automate descriptions. I wanted to personalize them too.

Content descriptions were handwritten and tedious to write, so leadership wanted to automate the process with AI. My idea was to take it further and write a personalized reason to watch for every viewer.

BEFORE what it's abouthandwritten,same for everyone AFTER + why you should watchAI adds apersonal reason EXPECTED RESULT ↑ play ratemore peoplepress play

Choosing the right data.

When I design systems, the first thing I do is look at the data we have. Most of the time, we don’t need all of it, just the strongest signals.

From just watch history, I could infer a viewer’s favorite themes, tone, cast and creators, genres, and whether they like critically acclaimed titles. Titles they liked or hearted weighed more in their preferences.

DATA WE HAD WHAT I INFERRED watch history likes ♡(optional) completion ratetime of dayprofile details themes tone cast + creators genres critically acclaimed I chose the strongest signals and inferred the rest.

I started with the smallest version to prove my personalization idea: just watch history and likes. Fewer inputs meant a lighter system to build and maintain, so we could ship sooner.

03 — Design

Four agents each have one job.

I started with one agent doing everything, but quickly hit a wall. The more tasks it juggled, the less personal the descriptions got. So I split the workload, adding agents one at a time to fix each specific problem.

The first agent I added was a pattern analyst. It studied each viewer’s watch history and became an expert on their preferences. My original agent became the writer, now free to focus on tone and voice.

Then I added two checks. An editor reviewed each draft against UX writing and legal guidelines from Warner Bros. and partner studios like A24 and Studio Ghibli. A critic compared three drafts and picked the most personalized one, using an LLM-as-a-judge approach. It told the writer why each draft won or lost, creating a feedback loop inside the system.

The result was a chain of four agents, built with Gemini and Codex.

pattern analyst finds patterns inwatch history(crawls IMDb,Wikipedia, OMDb) writer writes 3 options,135 characters each editor checks UX writing+ legal guidelines critic picks the mostpersonalized (wouldonly work for thisviewer) why each draft won or lost

See my agents in action!

Meet a Friends fan.

WATCH HISTORYFriends ♡The O.C. ♡Lady BirdBig Little Lies RECOMMENDED TITLE Beat Bobby Flay pattern analyst big personalities,competitive streak,Friends is afavorite writer 3 drafts,135 characterseach editor checks the rules:no spoilers,not an ad critic picks the mostpersonalized Before: “Talented chefs battle for the chance to take down Chef Bobby Flay.” “As if Monica Geller from Friends got her own cooking show.All her sassy, competitive fire in a delicious 20-minute duel.”

Same show, a different reason for every viewer.

At scale, my system gives millions of viewers their own personalized description.

WATCH HISTORYPERSONALIZED DESCRIPTION House of the Dragon+ my AI system a Lord of the Rings fan “Love the epic scale of TheLord of the Rings? Swap thefellowship for rivalfamilies in a brutal fightfor the Iron Throne.” a Big Little Lies fan “Another iconic femalerivalry. Two childhoodfriends split over athrone, and their feud setsa kingdom on fire.” a White Lotus fan “The rich and ruthless, nowwith dragons. Messy familydrama and a fight for thethrone that turns into all-out war.”

04 — Iteration

Six weeks of reviews to perfect the rules.

Every week I revised agent prompts and generated new descriptions to review with our UX writer and a senior designer. Instead of relying on just my judgment, we aligned on what a good description sounds like.

write agent prompts generate personalizeddescriptions review with UX writer+ senior designer iterate × 6 WEEKS

I translated what we agreed on into rules every agent follows.

what it mentions how it sounds what it never does ✓preferred themes and tones ✓only titles they liked(with a strong connection score) ✓never more than 1 title ✓well-watched friend, not an ad ✓10th-grade reading level ✓no exclamation points ✗feel creepy or invade privacy (like revealing when or how you watch) ✗spoil the story ✗mention titles not in the HBO Max catalog

Gemini Pro was the only approved model, and in 2025 it was less capable than today’s models. Most gaps in output were because the agents didn’t have the data to make the judgment call. So I gave it to them.

Connection scores: The pattern analyst scored how strongly each liked title connected to the recommended one. A title was only name-dropped if its score passed a threshold I calibrated.

Prompt vocabulary: A model doesn’t understand words the same way a team does. In prompts, I called the descriptions “recommendations” instead of “personalized descriptions,” “synopses” or “blurbs.” The voice shifted from promotional or dry to a friend telling you what to watch, with a sense of humor.

Designing for AI means learning how it understands language, not just using your team’s internal vocabulary.

05 — Testing

I built a tool to test the descriptions blind.

I didn’t want AI to be the only judge of itself, so I added human evaluation on top of the critic agent.

To compare handwritten descriptions against the AI-generated ones fairly, I built an internal testing tool in React, TypeScript, and Firebase. Testers pick a show they’ve watched, then choose between two unlabeled descriptions. I worked with a UX researcher to design the study and word the questions so they wouldn’t lead testers. More than 20 teammates tried it.

Testers fill in their watch history, pick which descriptions they prefer, and rate them.

Testers said descriptions helped them decide.

  • 2blind rounds with 16 viewers
  • 54%test users said AI descriptions helped them decide whether to watch

We hadn’t launched, so this was our early-stage signal for play rate.

After round 1, I rebalanced plot and personalized reason to watch. Testers still expected the description to tell them what the show is about, since that was their existing mental model.

Earlier research showed 5 of 7 viewers left without pressing play. Here, more than half of testers said the description helped them decide, which is a promising start.

What testers told us:

  • Hilarious, fun to read, and fairly accurate.
  • Very me.

06 — Impact

Segmentation tested just as well and saved us millions in API costs.

Blind tests showed that writing for 17 user segments did as well as personalizing for each user. With about 128 million HBO Max users, that meant 17 descriptions per title instead of 128 million, saving the company an estimated millions of dollars in API costs.

  • 128Muser
    generations
  • 17user
    segments

The trade-off is that segmentation puts each viewer in one bucket, when in reality viewers like multiple themes. My 1:1 system avoided this by selecting 2–3 preferred themes per viewer, like female-led stories and mystery.

I presented the findings to the CPO and the VP of Design, then handed my work to the design team with the 17 recommended user segments so they could keep iterating.

If I did it again today.

New AI models come out every few months, and each one reads prompts differently. I’d now build an evaluation set first: a fixed list of inputs to run on the old and the new model. I’d adjust the prompts where results drop and rerun until the new model matches or beats the old one. This would future-proof the system.

My personalized descriptions helped 54% of test users decide whether to watch. I steered the team from 1:1 personalization to segmentation, which cut generation from ≈128M descriptions per title to 17, saving an estimated millions in API costs.