---
name: quiz-me
description: Use when a user asks to quiz, test, drill, or practice weak Japanese words from their Nihontekini decks.
---

## Workflow
1. Choose recorded vs informal mode. Recorded mode uses `make_quiz` (max 30 questions), which consumes one AI-generated quiz and saves a quiz record; avoid making a quiz just to preview it. Informal mode calls `get_learned_words`/`get_weak_words` and is not recorded.
2. If a user wants the weakest vocabulary, use `get_weak_words` to choose the level. If recorded mode is requested, call `make_quiz` once; handle empty decks, auth errors and `isError` refusals without retrying. If it refuses, suggest informal practice from existing words or Test me in the app.
3. Ask one question at a time. Do not reveal correct option, kana, or romaji before the user answers if those are under test. Keep correct answers private while asking.
4. Give short feedback after each response; store the user's selection as the exact requested `picked` representation. Don't record questions they didn't answer.
5. For recorded quizzes, call `record_quiz_answers` once using the exact returned quiz_id and card_ids, up to 30 responses. If the session ends early, record only answered questions when appropriate.
6. Give the score and 3–5 words to revisit. If no practice data exists, explain what to review first.

## Cross-workflow rules
- Use only Nihontekini MCP tools with exact server-returned IDs; never invent card, deck, quiz, or user IDs. Authentication failures: ask the user to reconnect the plugin.
- Card `meaning` must contain only plain English glosses (e.g. `["to study", "diligence"]`). Put kana in `reading`, usage and grammar in `explanation`, and never include romaji, pronunciation, kana, part of speech, or annotations inside `meaning`. The `reading` field contains kana, never romaji.
- Examples should be natural, with Japanese text in `jp` and, when available, Anki-style furigana in example `reading`, e.g. `日本語[にほんご]を 勉強[べんきょう]しています。`.
- In chat, when showing Japanese, give kana and romaji alongside kanji as appropriate; in a question that tests reading, never spoil the answer with reading or romaji.
- Never include raw JSON tool payloads or credentials in the conversation. Summarize results concisely in the user's language.
- Send at most 200 cards per add/suspend/delete request and 30 questions per quiz. On a 429 Too Many Requests, respect Retry-After and avoid hammering.
- If a tool refuses (`isError`), pass its message on plainly and stop without retrying unless the message says what to change. Never claim the assistant can buy or change anything on the account.
- Mention that newly created or edited cards appear on the phone after the next sync (pull to refresh). Never promise immediate propagation.