| name | anime-recommend |
| description | Use when the user wants anime recommendations, taste profile analysis, or asks "what should I watch" / "summarize my taste". Also triggers passively — record user taste signals during normal conversation.
|
anime-recommend
Passive: Record Taste Signals
During normal conversation, update anime_taste.md via memory tool when the user:
- Rates or reviews an anime ("孤独摇滚真好看", "这番太无聊了")
- Marks done/dropped/on-hold via bangumi/update_collection
- States a preference ("我喜欢日常番", "不看机甲")
Append only. Dedup. Move items between Liked/Disliked on attitude change.
Active: Recommend
Step 1 — Get all data (single call)
Call anime-recommend/default.
Script does: Fetches user collections, builds tiered confidence stats, fetches calendar, computes Bayesian weighted scores — all in one call.
Script output has two parts:
Part 1 — Taste Profile:
- If up-to-date: returns existing
anime_taste.md content.
- If new/changed: returns a tiered confidence report + titles list.
- Strong (≥50%) — genre/studio/director in ≥50% of liked titles
- Weak (30-50%) — in 30-50%
- Below 30%: already filtered out by the script
Part 2 — Scored Candidates:
- Calendar listing + Bayesian weighted ranking of all airing titles.
- Each candidate has: raw score, votes, weighted score.
weighted is for your internal ranking. raw is what users see.
Step 2 — Save profile (if new)
If the output says "ACTION REQUIRED", you must save the profile before recommending.
Call memory(action="write", filename="anime_taste.md"):
## Liked (Strong)
- genres: 战斗, 奇幻
- studios: MAPPA
## Liked (Weak)
- genres: 搞笑, 恋爱
## Disliked
- genres: 科幻
## Titles
- GIRLS BAND CRY → 10/10, loved: 作画, 音乐, 角色塑造
- 异世界舅舅 → dropped, disliked: 节奏太慢
CRITICAL: Only transcribe items that appear in the confidence report.
- If "Disliked Genres" section is empty in the report → omit
## Disliked entirely or leave genres blank.
- Do NOT infer preferences from the titles list. The titles list is context only — the confidence report is the source of truth.
- Do NOT copy genres from the example above — use the actual report output.
If profile is already up-to-date (output says "up-to-date"), skip this step.
Step 3 — Pick and present
From Part 2's weighted ranking, pick top 3-5. If taste profile exists, re-rank:
- Strong genre match → boost
- Weak genre match → mild boost
- Disliked genre match → penalty
Show raw score to user, NOT weighted score. Present each pick with: title, raw score, vote count, and why it matches (cite specific preferences). Flag disliked-genre overlaps as caveats. Skip titles already in user's collection.
No profile → present top weighted picks, note it's based on community ratings only.
