ChatGPT and Gemini in QA training: the experience of a Stepik course author for beginners

Artem Rusov, author of testing courses on Stepik and Udemy, shared how he integrates AI into teaching beginners. Based on his experience, models don't replace traditional courses but expand them: they help build custom learning environments, generate assignments, and provide students with more engaging practice.

Custom

Learning Sandbox A turning point for the author came in 2023, when without deep programming background, he built a complete sandbox for his main course using ChatGPT. This isn't just a collection of mocks, but a full system with frontend, backend, documentation, wireframes, and database. For a manual testing instructor, this turned out to be more than just an impressive demonstration of AI capabilities—it was a way to create an environment that perfectly aligns with course logic and isn't dependent on external limitations.

This approach quickly revealed practical benefits. An instructor can embed necessary bugs, quickly fix found issues, and build scenarios that truly test skills rather than the ability to complete a specific trainer. According to Rusov, students work not in a sterile learning wrapper but in a system resembling a real product.

An additional bonus for the course author himself—growth in personal expertise: when you build an educational service yourself, you understand code, infrastructure, and deployment processes much better.

Task

Generation and RAG The next stage is moving AI from course preparation into the actual user experience. Rusov is now developing a new version of trainers in Gemini, where students can generate unlimited new assignments in the same interface and according to the same template. This is no longer a story about a one-off experiment: to make everything work with high request volumes, you need to think about API keys, cloud platforms, and system resilience under load.

• Custom sandbox aligned with course curriculum Browser-based trainers for requirements, SQL, DevTools, and error detection Assignments with LLM-based grading and automatic feedback RAG search across a curated knowledge base instead of random generation Generation of illustrations and narrative framing for the course Meanwhile, the author is mastering RAG so students find answers not in arbitrary model outputs but in a pre-selected knowledge base. Following the same logic, the course now includes assignments where AI evaluates detailed answers, assigns scores, and provides comments. Such a tool, developed with a colleague, is embedded into an educational platform via admin panel and iframe. But Rusov immediately clarifies the boundaries of the approach.

"In education, it's important not to cause harm."

Gamification

Without Blind Faith Another direction is the visual and narrative packaging of the course. The author shares how he transformed his longtime mascot, an artichoke, into a character named Arty: one who travels through ancient civilizations and tests APIs. For this setting, he collects separate sandboxes and images generated through Nano Banana.

The idea isn't decoration for decoration's sake, but making the course more memorable and looking like a cohesive product rather than a collection of disconnected lessons and tests. At the same time, Rusov doesn't oppose AI to traditional learning formats. Regular tests, matching assignments, timeline reconstruction, and fill-in-the-blank exercises haven't gone anywhere, and all texts created with models are carefully reviewed, edited, and fact-checked manually.

His main thesis is simple: you can't mass-generate "AI slop" because the author is responsible for every fragment of the course. In this scheme, AI isn't a replacement for methodology but a tool that works only with clear boundaries and editorial oversight.

What

This Means Rusov's case shows that AI in online education is beginning to deliver value not at the level of slogans but in specific product elements: from sandboxes and answer grading to personalized trainers and visual layers. But the advantage emerges only where the course author maintains responsibility for content, quality, and educational logic rather than outsourcing everything to generation.