Designing personalized Reasons to Watch for HBO Max

Reasons to Watch appears when a viewer focuses on a title. Instead of repeating the synopsis, it argues for the title. Internal research had already shown the pattern helped: it lifted clarity around why something was being recommended by 4%.

The scalable version of that idea assigned every viewer to one of 17 affinity groups and wrote one piece of copy per group. It was practical for cost and for a catalog this size. It also flattened taste. The same person can want a prestige drama, a competitive cooking show and a comfort sitcom in one week, for three completely unrelated reasons.

The brief I set myself: write for one viewer rather than their segment, without breaking the 135-character limit, Max's voice, or the spoiler rules.

Two signals were available, and they are not interchangeable. Watch history is broad and behavioral — it shows what someone puts on. Explicit likes are narrow and endorsed — they show what someone would defend. The first model I built used history alone, and it name-dropped titles a viewer had merely watched, sometimes titles they had abandoned halfway.

That version tested worse than the generic copy. Not because the writing was weaker, but because a wrong name-drop is louder than no name-drop. It reads as a system claiming to know you and getting it wrong.

Personalization was never the name-drop. It was the explanation of fit.

Splitting the work was not about accuracy. It was about diagnosis. Once each stage returned a structured handoff, a weak sentence had an address: a bad connection is an analysis problem, flat language is a writing problem, a broken character count is an editing problem. Each one gets tuned without disturbing the others.

The editor is the design system

Most of what makes the copy sound like Max lives in one agent. These are product rules, not style preferences — several exist because legal or brand would reject the alternative, and every one of them started as something the model kept doing wrong.

• Length135 characters, after editing
• NamingNever "this show", never the title itself
• AwardsNo Emmy, Golden Globe or HBO — use "critically acclaimed"
• ReferencesFull title, italicised, one maximum
• PunctuationNo exclamation points. No em dashes.
• PresumptionNo "If you enjoy…", no "You love…"
• RhythmNo three-adjective stacks. No one-to-three word sentences.
• Banned words — with one exceptionNever "grit" or "chess". Never "delicious" unless the title is genuinely culinary.

That last exception is the one I point at when someone asks whether the rules are real. The pitch for Selena + Chef ends "…all her sassy, competitive fire in a delicious 20-minute duel." The word survives editing because the show is about cooking. A rule with a working exception is a spec; a rule without one is a filter.

Everything above could be true and the feature could still not be worth shipping. Personalized copy will always beat generic copy when the reader knows which is which. The only honest question was whether it beat the default when nobody knew which was which.

I worked with design technologist Travis Swan to build an internal evaluation tool. Participants entered their own watched and liked titles, then compared default copy against generated copy blind. After each choice the tool asked why — plot and theme balance, a familiar connection, attention, or a sense that the writer understood the title.

The tool took weeks to build, and I did not want to wait weeks for signal. While it was in development I ran the same comparison asynchronously in a spreadsheet — uglier, slower to analyze, and it answered the question a month earlier.

What the prompts lost along the way

The agent guide keeps every rejected phrasing, because each one failed for a reason worth remembering.

• "You are a blurb writer"

  Accurate and impersonal. Produced catalog copy.

• "You are a friend of the user"

  Personal, and it started using slang. Max does not use slang.

• "You are a friend who is really good at making recommendations"

  Chosen. Warmth without the register slipping.

• Critic scores "Personalized" and "Emotional Appeal"

  Killed. The writer produced three near-identical drafts, so the critic had nothing to choose between. Fixed upstream instead, by demanding three distinct angles.

• Longer, more detailed agent prompts

  Killed. Past a certain length the copy reads as constrained — it sounds like something obeying rules.

Qualitative feedback landed on three strengths, consistently: the copy matched a mood the viewer was already seeking, it was genuinely enjoyable to read, and the connection to a past favorite made the recommendation easier to believe. The work went on to inform the team's exploration of personalized copy in single-content promotions.

If it shipped, three metrics would settle it. I would build them with engineering before launch rather than after, since none of them can be reconstructed from a copy test.

The third is the one I would watch hardest: the share of sessions that end without playback. It is the inverse of the first, and it is the only one that catches a reason that persuades someone to open a title and then disappoints them.