How recommendations are made
Last updated October 9, 2026
Palate turns your Letterboxd export into a taste report and a list of films to watch next. Code does the counting. An AI model, Claude by Anthropic, is used only for judgment: naming the kinds of film you like and choosing films that fit them. This page explains each step.
1. Reading your export
Your export ZIP is opened in your browser. We receive a list of films: title, year, your rating, and whether you watched, liked, favourited, rewatched or watchlisted each one. We also receive up to 20 short review excerpts, used only to name your taste and never stored. The privacy policy lists every field.
2. Matching films to TMDB
Each film is matched to The Movie Database (TMDB) by title and year. Letterboxd sometimes dates a film to its festival premiere and TMDB to its general release, so we also try one year either side. We only accept a match when the title agrees. We never request pages from Letterboxd.
For every film you watched, rated or not, we fetch its genres, directors, cast, keywords, countries, language and runtime.
3. Your statistics are counted, not guessed
- Counts of films by genre, director, actor, country, language, decade and runtime include every watched film.
- Averages use your ratings only. A director or genre needs at least 3 rated films before it appears in a highest-rated list.
- Several Letterboxd links can point to one film, such as diary entries. We merge them, so each film counts once.
- All charts and tables are on the My Stats page.
- Sub-genres, topics, story settings, historical periods and source works come from Wikidata, the free knowledge base (CC0). We look films up by their TMDB ID and send Wikidata nothing else.
4. Naming your tastes
Claude reads up to 120 of your films with the strongest preference evidence: favourites, likes, high ratings and rewatches. It also sees some lower-rated films, so it can tell what you watch often from what you love, and genre, director, country and theme counts from your rated films. It then names 4 to 6 tastes, or 2 to 3 if you have fewer than 40 ratings. Below 10 ratings we skip this step.
Each taste must cite films from your history as evidence, and its themes must come from those films. We check this in code. If a check fails, Claude gets one chance to correct its answer.
If you are signed in, Claude also sees up to 20 of the films you most recently dismissed: “Not interested” in a report or “Not for me” in an email. A dismissed film is never evidence for a taste. When several dismissed films share what a taste would describe, Claude narrows that taste to leave them out.
Films you logged without a rating do not shape your tastes or picks by their details. Their genres, people and themes are never used. If you liked, favourited or rewatched an unrated film, Claude sees its title as a signal of what you enjoy. Every unrated film counts in your statistics, and we never recommend it, because you have already seen it.
5. Gathering candidate films
We collect unwatched films from three TMDB sources:
- TMDB’s recommended and similar films for your 30 highest-rated films.
- Other films by directors whose films you rate highly or often.
- Well-reviewed films that share the genres and themes of each taste.
We remove every film in your export, your watchlist, and any film you marked “Not interested”. A candidate needs at least 50 TMDB votes (500 for the third source) and a TMDB audience score of at least 4 out of 10. The score floor is low on purpose: it removes films that most viewers disliked, not films that divide opinion.
Each taste then gets its own list of up to 40 films, ranked by fit: shared themes and genres, a director from the taste’s evidence films, and how many of your top-rated films pointed to it. A third of each list is kept for films found through your own ratings and directors, so widely acclaimed films do not take every place.
6. Choosing the picks
For each taste, Claude chooses 6 films from that taste’s list. It sees the taste, related films from your history with your ratings, and each candidate’s genres, themes, director, runtime, summary and TMDB score. It treats the score as a weak signal: when two films fit equally well, it prefers the better-received one. It also sees the films you dismissed, and avoids candidates that closely resemble several of them.
Before a pick reaches your report, code checks that:
- the film came from that taste’s list, so Claude cannot invent a film or recommend one you have seen;
- the reason names at least one film from your history, so every pick is explained by your own viewing.
A film can fit two tastes. When that happens, it usually appears under one taste only, so you see more different films.
7. Browsing the full list
Your report shows the top picks. The browse page shows every candidate from step 5, with its taste, TMDB genres and runtime, so you can sort and filter beyond the picks.
What this cannot know
- Only your export: if you saw a film but never logged it, we may recommend it.
- Taste names are an interpretation of your history, not a fact about you.
- TMDB scores come from TMDB users, not from Letterboxd members or critics.
Questions or ideas? Write to hello@palate.film.