Skip to content
4 Nov 2026 · AI tutoring · 2 min

How our AI tutor avoids giving away the answer

Md Nazmul Hosan ·

Four layers stop a nursing-exam tutor from leaking answers or giving unsafe advice. None of them is a clever prompt.

How our AI tutor avoids giving away the answer

Short answer: remove the answers from what the tutor can retrieve, let it act only at defined moments, check every reply before the student sees it, and test for leaks before every change.

Students preparing for nursing licensure exams practise thousands of questions. A tutor that tells them the answer is worse than no tutor: it feels helpful and removes the thinking the practice is for. When we built an AI tutor for a US nursing exam-prep company, the first requirement was simple to state and hard to meet. The tutor must not reveal, confirm or deny the correct answer during practice.

Prompt instructions alone do not meet that requirement. Models follow instructions most of the time, and most of the time is not good enough when a student can ask the same question ten ways. So we built four layers.

1. Remove the answer before indexing

The tutor retrieves passages from the course content to ground its coaching. If those passages contain the answer key, the model can find it. We remove answers from the content before it is indexed. The tutor cannot leak what it cannot see.

2. Coach only at defined moments

The tutor acts at three moments: a hint before the student submits, a Socratic chat before submit, and a reflection after submit. Each has its own prompt and rules. Before submit, every turn ends with a guiding question. After submit, the reflection explains what the student got right, where the gap was and whether their confidence matched the result. A confident wrong answer is treated as the most dangerous pattern, because in clinical practice it is.

3. Check every reply before the student sees it

Every reply passes checks for answer leaks, unsafe clinical advice and basic quality. If a reply fails, the student sees a safe fallback instead. A fallback is less impressive than a good answer and far better than a leaked one.

4. Measure it

We keep an evaluation set of questions designed to trick the tutor into revealing answers, alongside normal coaching questions. It runs before any prompt or model change. If the leak rate rises, the change does not ship.

What it costs to run

Every AI call is logged with the model, tokens, estimated cost and latency. Routing tasks to the right model tier and caching repeat retrievals keep the cost per student predictable. A provider chain with automatic fallback means an outage at one provider does not take the tutor down.

Checklist for EdTech teams

  • Decide what the tutor must never do, then enforce it outside the prompt.
  • Ground the tutor in your own content, and clean that content first.
  • Build the evaluation set before the tutor, not after.
  • Log cost per call from day one.

Read the AI Exam-Prep Platform case study or see our AI tutoring service.

Share

Nazmul founded Door Soft in Dhaka and leads system design and delivery. He writes about building AI that works in production for mid-size companies, and about the products Door Soft runs itself.

About the author


All posts

← Back to the journey