Yandex Practicum showed where AI speeds up English course creation and where instructional designers are still needed

Yandex Praktikum shared how it uses neural networks when creating English language courses and where automation really saves time. The team's main conclusion is simple: AI is already taking some of the routine work off human hands, but quality educational content still cannot be created without specialists.

Where AI accelerates work

At Praktikum, AI was integrated alongside methodologists, not replacing them. A course still starts with design: the team defines skills, vocabulary and grammar by level, breaks them down by modules and lessons, and then edits the materials many times over. Neural networks are needed in this process where there is a lot of repetitive work and where a rough draft has low cost. They help to start faster, overcome the blank page effect, and scale the number of exercises to meet students' requests, especially those who want extensive practice on specific topics.

• Draft educational texts for specified vocabulary and level Standard exercises for grammar and vocabulary Generation of some illustrations for children's courses * Voice synthesis of materials using synthesized voices According to the team's observations, AI works best with text and structured tasks. For example, it can prepare gap fill, drop-down, matching or put in order exercises if given a format and boundaries in advance. This provides significant time savings on repetitive tasks. However, even in such scenarios, the result does not go directly into the course: it is checked, edited and integrated into the overall program logic by people.

Where the model fails The most painful area is the quality of pedagogical logic.

The model can produce a grammatically plausible exercise, but poorly understands which incorrect options are actually useful for a student. A good distractor is not just a wrong answer, but a typical mistake made by a Russian-speaking student at a particular level. AI often suggests options that are too obvious or monotonous, so the exercise looks neat but does not develop the skill well.

If a methodologist has to manually fix all the weak points, the time savings quickly disappear. There is also a deeper problem: the model does not feel real communication and does not know the context of the entire educational program. It can generate an example with Present Perfect that is formally correct but sounds unnatural in living speech.

Or insert a word that exceeds the student's level and becomes a blocker for the task. This is also where cultural context errors arise: in one example, a topic about a house moved into a tree house, which is normal for English-speaking environments but is not obvious to a Russian-speaking student. Similar limitations are visible in visuals.

In children's courses, 80-90% of illustrations are already generated by neural networks when dealing with simple objects and characters. But as soon as a complex plot is needed, precise quantity of objects, numbers on cards, or an unusual character with strictly defined details, generation starts to fail. With sound, the situation is better: speech synthesis already provides natural voices, emotions and accents, but materials are still listened through again and sent for re-voicing if the pace, accent, or voice distinctiveness do not fit.

How they achieve quality

At Praktikum, they emphasize that good results come not from one successful prompt, but from a long series of iterations. The team showed an example: in the first version of the prompt, the model violated the task format and even substituted the target vocabulary where the student was supposed to recover it themselves. After that, the prompt was refined with precise audience, topic, number of sentences, answer structure and format constraints — and quality improved noticeably.

But even after this, manual review was necessary. Over a year and a half of working with generative models, the team developed a set of practical rules that significantly improve draft quality in real work. This is not about "magical" communication style with the model, but about discipline in task specification: you need to show an example, strictly fix the format, length and number of elements, and remove everything that blurs the target grammar or vocabulary.

Exactly such constraints have the strongest impact on the usefulness of results.

• Provide the model with an example of the desired format instead of relying on guessing Run new tasks in a clean context so that old requests do not affect the answer Specify exact length, number of elements and parts of speech for target words Explicitly limit unnecessary grammatical constructions if you need one specific topic Do not rely on the model's self-checking as a guarantee of quality There is also an interesting observation about prompt engineering: a preamble in the spirit of "you are a qualified specialist" almost did not change the output. Examples, constraints and specificity worked much more effectively. For teams building educational content on generative models, this is a useful insight: there is no magic formula, and stable quality is still assembled manually from review, iteration and subject matter expertise at each step. In other words, value comes not from the role but from task specification.

What this means

The Yandex Praktikum case demonstrates well the current place of AI in education: it is a powerful accelerator for drafts, routine exercises, some visuals and voice-over, but not an independent course author. For self-learning, neural networks are useful as a helper for practice, explanations and quick feedback. But program design, selection of quality tasks, level control and protection against errors still remain a human task in almost all serious scenarios.