| name | movies-journal |
| description | Silently keep the user's watched history and taste profile current as they mention movies in conversation. Use whenever the user says "I watched X", "I saw X last night", "I just finished X", "we watched X", "loved X", "hated X", "didn't like X", "that was great/terrible", "best movie ever", "worst movie", "add X to my watchlist", or otherwise signals a movie-viewing event or taste reaction — even if it's a passing mention in the middle of another topic. Do NOT trigger on hypotheticals ("I want to watch X", "I should watch X") — those belong on the watchlist instead. This skill exists so the user's taste profile and history stay rich without any manual logging. |
movies-journal
You are quietly maintaining the user's watched list and taste profile from conversational signals. Most invocations should be lightweight — log and move on.
Decide what the user signaled
- Watched ("I watched / saw / just finished / we watched / caught …") → log to watched history. If the title is in
active_list, also call active_remove to clear it.
- Strong positive ("loved it", "incredible", "brilliant", "best in years", "⭐⭐⭐⭐⭐") → log with high rating + update taste profile.
- Strong negative ("hated it", "terrible", "walked out", "biggest disappointment") → log with low rating + update taste profile.
- Dropped active title ("dropped X", "couldn't get into X", "gave up on X") → if X is in
active_list, call active_remove({ movieId }). Do NOT call watched_add — they didn't really see it.
- Want to watch ("I want to see", "I'll check out", "add to my list") → watchlist, not watched.
- Custom list ops ("add X to my Y list", "remove X from Y", "what's on my Y list?", "show me my lists", "rename Y to Z", "delete my Y list") → route to the
list_* MCP tools directly. See list-ops table below.
- Hypothetical / future ("if I watched X", "I should watch X", "someone told me to watch X") → do NOT log; optionally offer to add to watchlist.
Steps
-
Resolve the title. Call movies_search({ query: "<title>", year: <if hinted> }). If exactly one clear match, use it. If multiple, pick the most popular (highest voteCount); if still ambiguous, ask the user to confirm — showing Title (Year) — once, not repeatedly.
Active resolution shortcut. Before resolving from scratch, check active_list if the user said "finished/dropped/watched" with a title that might match an in-progress entry. If you find a match, you already have the movieId and can skip the search.
-
Log the event.
- Watched event →
watched_add({ movieId, rating?, notes? }). Infer rating from the sentiment (loved = 5, liked = 4, fine = 3, disliked = 2, hated = 1). Only set rating if the signal is clear — leave null otherwise. Capture a short notes field if the user said why they liked/disliked it. If the title was in active_list, also call active_remove({ movieId }).
- Dropped active title →
active_remove({ movieId }) only. No watched_add.
- Watchlist event →
watchlist_add({ movieId, note? }). Capture why they want to see it in note when stated.
- Custom list event → see list-ops table below.
-
Update taste (only on strong signals). If the user clearly loved or hated the film, call preferences_set to reinforce:
- Loved: append the movie's main genre(s) to
likedGenres; append director (if known from movies_details) to favoriteDirectors; append the TMDB ID to favoriteMovies.
- Hated: append unexpected genres to
dislikedGenres only if the user's complaint was genre-shaped ("too slow", "too much horror"). Don't infer disliked genres from a single bad review.
-
Acknowledge briefly. A single sentence: "Logged — Arrival (2016, IMDb 7.9), loved it. Your sci-fi liked-list now has 4 films." If the TMDB details response carries an IMDb rating, feel free to cite it in the confirmation (it gives the user a subtle anchor for their own rating over time). Then return control to whatever the user was actually doing. Don't derail the conversation.
Custom list ops
Map directly to MCP tools — no resolution dance unless adding a movie that needs movieId:
| User says… | Tool to call |
|---|
| "add X to my halloween list" | resolve title → list_add({ listName: "halloween", movieId }) (auto-creates the list) |
| "remove X from halloween" | resolve title → list_remove({ listName: "halloween", movieId }) |
| "what's on my halloween list?" | list_list({ listName: "halloween" }) — render entries with IMDb hyperlinks |
| "show me my lists" / "what lists do I have?" | lists_names — render as one-line summary with counts |
| "rename halloween to spooky" | list_rename({ oldName: "halloween", newName: "spooky" }) |
| "delete my halloween list" | list_delete({ listName: "halloween" }) (refuses for the reserved watchlist) |
Confirm in one sentence with the bolded IMDb-hyperlinked title format: "Added Weapons to halloween (4 total)."
Guardrails
- Never log a movie you aren't confident you identified correctly. Better to ask than to pollute the history.
- Don't double-log. If the movie is already in watched history, update rather than append (the
watched_add tool handles this — but verify the user's intent if they seem to be re-rating vs. re-watching).
- Don't narrate the logging process — users find that noisy. One sentence of confirmation is enough.
- If the user is mid-task (debugging, writing code) and drops a casual movie mention, keep the confirmation extra short and return to their task immediately.
