AI tutors are a popular topic. They offer the opportunity of 24/7 personal learning support for students, which is particularly beneficial in universities with a large number of students participating online. The game SQL Scrolls currently features a leaderboard and an AI-based task recommendation system, but does not include an AI tutor. Currently, players who do not know the answer can display the model answer, but get no personalised help.

After three failed attempts and upon request, a Large Language Model (LLM) generates a personalised hint, based on the model answer and the player's input. The prompt instructs the LLM to adhere to pedagogical guidelines while revealing part of the correct query and explaining it. We developed an algorithm to generate SQL skeletons and embed them within the task data to facilitate AI feedback. We established an automated quality control system to evaluate outputs from several LLM models, including models from OpenAI and Google.

New model versions of Google and OpenAI, Gemini-2.5-flash and GPT-4.1, both deliver accurate instructive hints. The final version of the AI tutor features Gemini-2.5 Flash, since it is available in a free tier. Gemini's thinking feature is disabled, allowing the model to provide hints in under one second, while the answer quality remains high. The hints often contain a skeleton of the query and enable players to make progress without revealing the complete solution. The literature shows that receiving a hint instead of the model solution is beneficial to the learning process. The study demonstrates the potential of AI-based tutors tailored to a clearly defined domain and having a clear purpose. Such tutors could also be used in other courses or contexts, offering students economically viable, personalised support around the clock.