You already have recommendation algorithms. Netflix, Spotify, YouTube, and every bookstore on the internet are optimizing for what you'll click next. The problem is that these systems know what you watched but not why you watched it, and they optimize for engagement rather than satisfaction, which is how you end up trapped in a filter bubble of content that's fine but never surprising.
Michael Tiffany
Netflix says that over 80% of what its members watch comes from algorithmic recommendations, and yet the most common experience on the platform is scrolling for twenty minutes, settling for something mediocre, and falling asleep on the couch. Spotify builds you a Discover Weekly playlist that's eerily accurate about your taste in tempo and key signature but couldn't tell you why you've listened to the same Radiohead album fourteen times this year. The algorithms know what you consume; they have no idea why you consume it, which means they can predict the next thing you'll click on but not the next thing that will genuinely move you, challenge you, or make you call a friend to say "you have to watch this."
This is the gap your AI agent can fill. The method is different from any of the previous articles in this series because you're not starting from scratch, you already have years of behavioral data scattered across platforms. The work here is not capturing new reactions (though that helps), it's translating the preferences you've already demonstrated into language rich enough for your agent to understand why you like what you like, so it can reason about what you'd like next rather than pattern-matching against what you've already seen.
Scientific American published a useful overview of how recommendation systems work and where they break down. The short version is that most systems use collaborative filtering (people who watched X also watched Y) and content-based filtering (if you liked this thriller, here's another thriller) in combination, and both approaches converge on a well-documented problem called the filter bubble: the more you watch, the narrower your recommendations become, because the algorithm reinforces your demonstrated preferences and gradually stops showing you anything outside them.
The deeper issue is that these systems optimize for engagement metrics: clicks, watch time, completion rate. They don't optimize for the thing you actually care about, which is whether the experience was worth your time. You might finish a mediocre series out of momentum and mild curiosity about the ending, and the algorithm reads that completion as a strong positive signal, which means it will recommend you more shows with the same narrative structure, pacing, and emotional register, burying you deeper in a groove that felt like a rut even before you finished the last one.
Your agent can escape this trap because it doesn't optimize for engagement. It optimizes for whatever you tell it to optimize for, which means you can teach it the difference between "I watched the whole thing" and "I loved it," a distinction that no behavioral signal can reliably capture.
The fastest way to teach your agent about your entertainment preferences is to contrast what you loved with what you quit, because the reasons you abandon something are often more specific and more instructive than the reasons you finish it.
Pick five things you genuinely loved across any medium: a movie, a show, an album, a book, a podcast, a game. For each, tell your agent what made it special, going beyond genre and plot into the qualities that hooked you. "I loved Severance because the pacing was deliberate without being slow, the world-building was revealed through behavior rather than exposition, and the tone sat in a narrow band between dread and dark comedy that almost no other show occupies. Also, the production design was immaculate; the office interiors felt oppressive in a way that was clearly intentional and consistently maintained." Your agent now knows you care about tonal precision, visual design as narrative device, and patient world-building, and that genre is secondary to execution.
Now pick five things you started and abandoned. "I quit The Witcher after three episodes because the timeline jumps were disorienting without being rewarding, the dialogue was expository in a way that assumed I hadn't figured things out from the visuals, and the fight choreography was impressive but the emotional stakes were so low that the action felt decorative." Your agent now knows that nonlinear timelines don't bother you in principle (you loved Severance, which plays with time), but they need to serve comprehension rather than complicate it, and that action without emotional stakes is a dealbreaker regardless of production value.
The contrast between the love list and the abandon list is where the signal lives. If you tell your agent you loved five slow, atmospheric pieces and abandoned five fast-paced, plot-driven ones, it learns something about your pacing tolerance. If you loved three things with morally ambiguous protagonists and abandoned three with clear heroes and villains, it learns about your appetite for moral complexity. These patterns are invisible to Netflix's algorithm because it never asked you why you stopped watching; it only registered that you did.
Entertainment preferences shift with context in ways that no platform currently tracks. You watch different things when you're alone than when you're with your partner. You listen to different music when you're working than when you're cooking dinner. You read different books on vacation than during a stressful work week. The same person who wants a three-hour Tarkovsky film on a quiet Sunday evening wants a ninety-minute action movie on a Friday night after a brutal week, and those aren't contradictions; they're context-dependent rules.
Tell your agent about these contexts: "When I'm watching alone on a weeknight, I want something under an hour, visually interesting, and tonally consistent. When my partner and I watch together on weekends, we need something we can both enjoy, which usually means strong ensemble casts, some humor, and nothing too grim. When I'm on a plane, I want something absorbing enough that I forget I'm in a middle seat but not so complex that turbulence ruins it."
These context tags transform your agent from a single recommendation engine into something more like a librarian who knows your moods, one who wouldn't suggest the same book for a beach vacation and a hospital waiting room.
Unlike the closet audit (wardrobe) or the product audit (skincare), entertainment preferences benefit from continuous, low-effort feedback because you're consuming new content regularly. The simplest habit is a one-sentence reaction after you finish something, or after you quit:
"Finished Shogun. Exceptional. The patient pacing I loved in Severance is here too, and the cultural specificity was so detailed it felt educational without being didactic. Add this to the reference list for what I consider great television."
"Tried a new podcast about the history of mathematics. The content is interesting but the host's vocal fry is so distracting I can't focus. I don't think I can get past it. File under: great subject, unusable delivery."
"Put on the new Beyoncé album while cooking. The country-adjacent tracks are surprisingly good for weeknight cooking energy; the slower ballads are not. I'd like to hear more music in that lane: rhythmic, warm, slightly genre-defiant."
Each of these takes fifteen seconds and teaches your agent something a recommendation algorithm never could: that vocal delivery matters as much as subject matter, that you use certain music for specific household activities, and that you maintain a mental reference list of great television against which you evaluate new shows.
Ask it to recommend one movie, one album, and one book for a specific context: "I have a solo Saturday afternoon with nothing planned and I'm in a contemplative mood." A good recommendation would draw from the patterns in your love list (patient pacing, tonal precision, moral complexity) while accounting for the context (solo, contemplative, ample time). A bad recommendation would suggest whatever is trending, or whatever is generically well-reviewed, or whatever your demographic tends to consume.
When the recommendation misses, the correction is the teaching: "You suggested a [well-known recent thriller]. I can see why based on my taste for dark premises, but I told you I prioritize atmospheric storytelling over plot mechanics, and everything I've read about this suggests it's plot-driven with a twist ending. Try again with that distinction in mind."
Should I import my watch history or listening history from streaming platforms? It's useful as background data but insufficient on its own, because platform histories contain everything you consumed without distinguishing between what you loved and what you tolerated. If you import a history, annotate the highlights and the disappointments; the raw list without your commentary is just noise.
What about books? Recommendation algorithms are notoriously bad for books. Book recommendations are an excellent use case for this approach, precisely because the existing algorithms are so weak. Telling your agent "I love books that teach me something about a world I know nothing about while maintaining a strong narrative voice, and I lose interest when the prose is workmanlike even if the story is compelling" gives it more to work with than a list of genres ever could.
How do I handle shared accounts or watching with other people? Specify the context. "My partner and I watched this together and we both liked it" is a different data point than "I watched this alone and loved it." If your agent knows who you watch with and what works in each configuration, it can recommend differently depending on whether you're asking for yourself or for a shared evening.
Won't my agent just become another filter bubble? Only if you let it. The advantage of a general-purpose agent over a recommendation algorithm is that you can explicitly ask it to surprise you: "Suggest something I wouldn't normally choose but that you think I'd appreciate based on the deeper patterns in my preferences." No streaming platform has a button for that.
Think of the last thing you watched, read, or listened to that you genuinely loved, and tell your agent what made it special in three or four sentences. Then think of the last thing you quit, and tell your agent why. That contrast is the seed of a preference model more useful than any algorithmic profile you've built over years of passive consumption.
