Turning the thing you just asked about into practice questions

Curiosity is where retention starts

You remember what you wanted to know. A topic you chased down because it bothered you sticks in a way that the next item on a revision list does not, and every study method worth the name is built on that difference.

Most tools throw it away. Reading happens in one place and questions happen in another, and moving between them costs enough friction that you do it later, or never. By the time you sit down to practise, the confusion that made the topic memorable has faded and the topic list came from somewhere else entirely.

The best moment to practise a topic is the moment you stopped being confused by it.

What the button does

A button reading Practice this topic sits under the chat composer. Press it and a modal opens saying Reading your conversation. Here is what that means, step by step.

• It reads the conversation, not the last message. The whole exchange goes in — up to the last thirty messages, capped at 800 characters each. Ask about a murmur, get sidetracked into valve anatomy, come back to the murmur: all of it counts.
• One bounded call maps it onto the bank. A small utility model reads the transcript and matches it against the question bank’s existing subject and system taxonomy. One call, nothing recursive, no browsing.
• It returns topics and a reason. One to three subjects, up to three systems, and a single sentence explaining why it chose them, shown in a tip box so you can disagree with it before anything is created.
• Then you are in the ordinary picker. Same screen you would reach from the question bank, with those subjects and systems already ticked.

The model names the topic. It never picks a question.

This is the part that decides whether the feature is trustworthy, so it is worth being precise about it.

The model does not return questions. It returns slugs — short identifiers for subjects and systems that already exist in the bank. Those slugs are matched server-side against the real taxonomy, and any the server does not recognise are dropped. Not corrected, not guessed at, dropped. The model cannot invent a subject that the bank does not have, because a subject it invents matches nothing and disappears.

Question selection happens afterwards and no model is involved in it at all. Once the subjects and systems are fixed, choosing which questions you actually see is plain policy over the filtered pool: deterministic, inspectable, the same rules every time.

The model is allowed to say what you were talking about. It is not allowed to decide what you practise.

The reasoning is the same as the one behind answers that carry their source: give the model the narrow job it is good at, and let verifiable machinery handle everything downstream of it.

You get the last word on the session

Pre-selected is not the same as chosen for you. Every subject and system arrives ticked and editable. Drop one that is not what you meant, add a system the conversation only brushed past, widen it if you want a harder mix.

• An optional question goal. Set a number if you are working to a target, or leave it open.
• An optional timer. Sixty, ninety or 120 seconds per question, for when the thing you need to practise is pace rather than recall.
• A session tied to its conversation. The session is created linked back to the chat it came from, so the questions and the discussion that prompted them stay connected.

Four ways to run it

The session modes are the normal ones, and they change which pool the questions come from:

• Practice. Fresh questions on the chosen topics.
• Unseen questions. Only questions you have never attempted in any session.
• Review mistakes. Questions whose latest attempt you got wrong, oldest first, so the ones you have avoided longest come back first.
• Bookmarked. Only the questions you starred.

A mode with nothing to serve is greyed out and tells you why, rather than starting an empty session and leaving you to work out what went wrong. Within a session, a question is never served to you twice.

What this does not do

It does not write questions. Everything you get is a board-style question that already existed in the bank, written and reviewed before you asked anything. The feature chooses where to look; it does not generate the material.

It does not read anything beyond the conversation you are in, and it does not claim to understand a conversation it could not place. Topics it cannot match are simply dropped, which means the honest failure here is a session narrower than you hoped, never a session about something you never mentioned. That trade is deliberate, and it is the same one described in our note on trusting AI in medical study.

The question bank covers Step 1. And practising a topic is not the same as being right about it: Med Guru is a study aid for exam preparation and revision, not medical advice. Verify against primary sources before anything you do clinically.