Recommendations didn’t understand why I liked something.
I used separate apps to track books, games, movies, and TV. Their recommendations often felt generic, and they rarely explained why I might like something. I wanted one place where I could see what the system thought my taste was, correct it, and have those corrections improve what it recommended next.
An editable model of personal taste
Tastemake looks at your favorites and ratings and proposes taste patterns across books, movies, TV, and games. Each pattern lists the titles behind it, anything that cuts against it, and how well supported it is.
You can confirm a pattern, reject it, say it’s only partly right, or limit it to certain contexts. Rejecting a pattern doesn’t erase the ratings behind it.
What corrections change today
What corrections don’t yet control
Saving a title you haven’t tried or marking it Not interested doesn’t count as evidence of taste. When AI is unavailable, recommendations fall back to catalog picks.
A more accurate citation wasn’t a better explanation.
A diagnostic trace turned up a mismatch. Tastemake’s explanations often pointed to a different rating than the one that actually decided the recommendation. So I tested a rule: every explanation had to cite the rating the scorer used to pick the title.
Trace alignment increased
Explanation quality decreased
For Breath of the Wild, the original explanation connected the game to Piranesi, and I rated that connection specific and true. With the rule in place, it compared the game to Return of the Obra Dinn on genre instead, and I rated that one incorrect.
The rule made explanations line up with the scoring trace far more often, but some of them got less meaningful. The deciding rating was often just a broad genre match, not the most useful reason to recommend something.
Decision: I didn’t ship the rule. I kept the diagnostic trace and turned my attention to how candidates get selected.
Better information wasn’t enough
Richer title descriptions looked promising in a held-out ranking test, but they didn’t meet the success criteria I’d set in advance. So I looked closely at one pattern that felt wrong to me.
- 01
Noticed
“Quirky character-driven comedy” lumped sketch and docu-reality shows like Nathan for You in with sitcoms like The Office.
- 02
Tested
I gave the model readable genres and short synopses. Its wording changed, but across 19 blind judgments from my own profile, none of the comedy pairings fit.
- 03
Understood
I’d answered “Partly” on that pattern. That kept its name in place but couldn’t say which part was wrong.
- 04
Explored
I prototyped a way to mark the words or titles that don’t fit and approve a suggested correction.
Research prototype only, not part of Tastemake. Seven trials I ran on myself, with hand-written suggestions standing in for the AI.
In one run, I marked five shows as fitting a different side of my taste. The suggested correction quietly dropped them and kept only the sitcoms, while its description still mentioned sketch and docu-style comedy. That’s when I asked, “Why does it only mention the sitcoms as what this comes from?” Every step had worked. It just hadn’t kept track of everything I’d told it.
Can understanding why someone liked something outperform genre matching?
Two bigger questions are still open: do people other than me find the recommendations useful, and does correcting a Taste Profile actually improve their next set? I also want Tastemake’s sense of someone’s taste to shape which titles it considers, not just how it orders and explains them.