What to Watch

Two Claude skills that recommend films and series from your watch history instead of from whatever is trending this week.

Every recommender you have used is pitching you the same twenty titles it pitches everyone. This one starts from the two thousand titles you already got through, the handful you cared enough to rate, and what you said about them afterwards.

Install

In Claude Code, as a plugin. This is the one to use, because it brings both skills plus the /what-to-watch-init and /watch-log-scan commands:

/plugin marketplace add jainulabudeenm/what-to-watch
/plugin install what-to-watch

From npm, which also covers Claude.ai and anywhere outside Claude Code. This installs both skills and the importers, without the slash commands:

npx @jainulabudeenm/what-to-watch

Add --project to install into the current repo instead of your home directory, so you can commit it and share it with a team.

Use

> what should I watch tonight

First run walks you through importing your history and asks five questions about your taste. After that it is instant.

It will ask you two to four questions back before it suggests anything. That is deliberate. "A thriller" is not enough information to pick on, and a recommendation made off a one liner reads as random.

> is Sinners worth it?

Checks your own library first. If you watched it in 2024 and gave it a 3, it will tell you that before it says anything else. That answer alone justifies the import.

> just watched Heat, the diner scene is the whole film, about a 9

Goes straight into your watch log, which is the strongest signal the recommender has.

Getting your history in

Three importers. All of them merge into one library, so you can run more than one.

MUST has no export button, so there is a puller for the public json API the app itself uses. No login, no key. Your MUST profile has to be public.

python3 ~/.claude/what-to-watch/pull-must.py <username>

Letterboxd: settings, import and export, export your data. Unzip it and point the importer at the CSV files. watched.csv, ratings.csv and watchlist.csv are the useful ones. Half star ratings are doubled onto a 1 to 10 scale.

Trakt: any export tool that gives you the API shaped json. Watched, watchlist and ratings files all work.

Then, whichever you used:

python3 ~/.claude/what-to-watch/import-history.py <files>

It writes ~/.claude/what-to-watch/library.json, keyed on title plus year so two sources give you one library rather than two half libraries. When sources disagree about how far you got, the one that says you got further wins.

None of the above? Skip the import. The skills still work off your taste profile alone and get sharper as the watch log fills up.

The two skills

what-to-watch does four jobs: pick something for tonight, vet a title you are unsure about, find something genuinely new, and tidy a wishlist that has grown into a graveyard.

watch-log captures what you thought. A tracking app records that you watched something. It cannot record that the second half fell apart, which is the part that makes the next pick better. It reads a notes folder if you keep one, proposes entries, and writes only what you approve. It never edits your notes and never surfaces anything in them that is not about a film.

Rules it will not bend

• Reasons on the film, not the cast. Never "from the director of", unless your profile names that person as a real signal.
• An unrated title means nothing. In most libraries under five percent of titles are rated. It will never read a blank as a dislike.
• No hype. Box office is not taste evidence.
• A guess never looks like a known quantity. Anything from outside your library is labelled as a suggestion.
• No spoilers.

Your data

Everything is in ~/.claude/what-to-watch/. Home directory rather than a project folder, so it follows you and cannot end up committed to a repo by accident. Plain json and plain markdown, yours to edit or delete.

Nothing is uploaded anywhere. The MUST puller reads a public profile, and the skill searches the web for genre and runtime on specific titles. That is the extent of it.

See taste.example.md for the depth of profile that produces good picks.

License

MIT