How to use AI to quiz you on your own notes

What should your AI study partner produce?

Start with five questions, a separate answer key and a short attempt log. Five is a manageable editorial starting point, not a research-backed optimum. Each question needs an ID, the skill it checks and a source location. The key should state what a sufficient answer includes and which alternatives are acceptable.

This guide is an original, untested approach for studying your own material. It adds a distinction that matters in practice: an incorrect answer and an incorrect question need different repairs.

The creator of DeutschLoop describes a German exercise that failed a learner even though the generated situation did not elicit the target word. The creator reports repairing the generator and undoing the bad record. That is a firsthand development account, not our test of the plugin. Its repository also labels several demonstrations as using a scripted learner; those histories do not establish a person's learning gains.

Bring one learning goal and readable source material

Choose a narrow goal such as explaining a concept, applying a rule to a new case, or remembering a procedure. Give the assistant the relevant notes, chapter excerpt or transcript you are allowed to use. Include a filename and section, page or timestamp so you can find each answer again.

Ask it to list the sections it could actually read. A link, blurry photo or filename is not evidence that the content was accessible. If your notes conflict with the course material, mark the disagreement and resolve it before using that item for assessment. A faithful question about an incorrect note can still teach the wrong thing.

• Your goal and starting level, plus one example of what you already understand.
• The selected source and version, with any topics to exclude.
• A short session budget and preferred format: spoken explanation, typed answer or worked example.
• A private place to keep the questions and attempt log. A plain document is enough.

Check the questions before you trust the score

Have the assistant show a small authoring sample with its answer key first. Reviewing the key is useful editing, but it exposes the answers. Label that round as practice with exposure; use fresh equivalent questions in a later session when you want to assess unaided recall.

Matthew Rhoads describes creating study activities from uploaded material and revisiting weak areas across sessions. This is his personal account, not an independent evaluation of our approach or a promise of improved results.

1. Open each cited source location. Confirm that it supports the answer, including qualifiers such as sometimes, unless or only after.
2. Check that the question gives enough context and tests the intended skill. Remove wording that reveals the answer or requires information absent from the source.
3. For an application question, list the reasoning the source permits. Do not make an invented detail part of the marking rule.
4. Allow equivalent correct wording. If more than one answer is defensible, revise the question or broaden the key before scoring anyone.
5. Keep uncertain items out of the scored set. Request a clearer source or an instructor's clarification when needed.

A worked example: repair the question, not the learner

This is a fictional film-club induction example. The entire supplied note, labeled Volunteer notes v1, section Arrival, says: Doors open at 6:30 p.m. The screening begins at 7 p.m. Late arrivals wait until the first break. The note does not give the time of that break.

Question A1: A guest arrives at 7:10 p.m. The screening is underway and no break has happened yet. What should happen before they enter? Answer key: They wait until the first break. Source: Volunteer notes v1, Arrival. Equivalent wording such as admit them at the first break is acceptable; an exact time is not required.

A flawed question asks: At what exact time can the 7:10 guest enter? If the agent invents 7:30 in the key, mark the item unsupported. Remove it from the scored set and correct it to A1. Keep a note explaining the change; do not record the learner's inability to give an unsupported time as a knowledge gap.

After that correction, ask a different case in a later session, such as a guest arriving at 7:20 while the screening is underway and no break has happened yet. It checks the same rule without simply repeating the original wording. Success on this tiny example is evidence about that attempt, not proof of broad competence.

Record what happened during the attempt

Ask one question at a time and wait for your answer. Request a hint when useful, but record that help. Then reveal the key and source, compare the reasoning and explain any difference. Treat the assistant's grade as a proposal you can challenge.

Anki's official studying manual likewise separates the question from the revealed answer and asks the learner to rate their response. Its ratings affect later review. The labels below are our suggested practice-log categories, not Anki's rating system or a validated grading rubric.

Keep a small record for the next session

Save the question ID and version, source location, date, your original answer, help used, the reviewed outcome and a next step. Keep attempted, assisted and invalid items distinguishable. A generated question is not a completed attempt; fixing its key is not a learning achievement.

For the fictional example, an entry might read: A1 v1; Arrival v1; October 10; answered they can enter immediately; hint used none; needs practice; reread the late-arrival rule. A second entry records the later answer separately. Do not overwrite the first attempt to make the history look better.

Agree on a later review you can realistically do. If you already use a spaced-repetition app, let its supported review system manage the schedule and preserve its existing history. Otherwise, keep a simple next-review date in your document. A universal interval or an AI mastery percentage is not required.

When the source changes, flag affected questions and version their keys before using them again. An old answer may have been correct under the old rule. Explain the change rather than retroactively treating it as a learner error.

Use the tools your assistant actually has

For Muse, Instinct, Dots or Grok Bot, start by checking whether your account can read the supplied material and return editable text. This guide does not establish compatibility with any particular account or integration. Without file access, paste a small permitted excerpt. Without durable storage, save the packet and attempt log yourself and attach them next time.

A reminder is optional. If you ask for one, confirm the saved schedule, timezone and destination, then check that a real run occurs. If scheduling is unavailable, use your own calendar. Do not require an app install or a new account just to try five questions.

A small brief to adapt for your next session

Help me practise [specific skill] from [source and section]. Confirm what you can read. Propose five questions with a separate source-linked key, acceptable answers and unresolved ambiguities. Let me review that sample; treat exposed answers as practice, not unaided recall. For a later quiz, ask fresh equivalent questions one at a time, wait for my response and record any hint or source lookup. Compare my answer with the source before proposing a result. Distinguish an invalid question from a mistake I made. Return an editable attempt log and a short next-review plan. Do not import a deck, change study history or schedule messages unless I ask.

Use the related prompts below for an overnight flashcard batch, a teach-back explanation or a short learning project. Add this guide's question checks and attempt record when you want to use those materials for practice.