How a lesson plan is made

A lesson plan in Lingara is a sequence of exercise sets built around a specific slice of vocabulary. You can generate one with a topic — ordering coffee, job interviews, talking to a landlord — or with no topic at all, in which case the level’s own vocabulary decides what the plan is about. This page is what happens in between.

What the model is given

Generation is a request to a large language model, and the request is not open-ended. It carries:

• the two languages — the one you speak and the one you are learning;
• the level, expressed as a difficulty descriptor rather than a bare number (see what a level means);
• your saved vocabulary and recent exercise history, so a plan does not hand you the same twelve words you cleared yesterday;
• the topic, if you typed one;
• whether pronunciation guides are needed for this language pair, and which romanisation system it uses.

That list is the whole of it, and it is published in full on our AI transparency page. The model is not given your email address, your subscription state, or anything you have said to the tutor.

What comes back, and what is checked

The model returns structured data, not prose: a plan title, a set of exercises, and for each exercise its prompt, its expected answer and its distractors. That structure is the first check. A reply that does not parse into the expected shape is not shown to you and is not stored; generation is retried rather than degraded, because half a lesson plan is worse than none.

Beyond the shape, the generator checks the things that are cheap to check and expensive to get wrong: that the vocabulary actually belongs to the level that was asked for, that exercises are not duplicated within the plan, and that the answer to a multiple-choice question is not also one of its distractors. These are not judgements about whether the lesson is good. They are the floor.

What no automated check can do is decide whether a sentence is natural, whether an idiom is current, or whether a register is right for the situation. That is the honest limit, and the app says so: content is labelled as AI-generated, and every plan carries disclosure metadata you can open. It is practice material, not an assessment, and it does not certify anything about your proficiency.

Why plans are pooled

Here is the part that surprises people, and it is worth stating plainly.

A finished lesson plan does not belong to you. It is shared into a pool, and other learners studying the same language at the same level are taught from that pool. Yours may be the plan somebody in another country works through tomorrow morning, and theirs may be the one you are handed tonight.

The pool exists for two reasons. The obvious one is cost and speed: generating a plan takes seconds and a model call, and serving a checked plan that already exists takes neither. The less obvious one is quality. A plan that has been through several learners has been through several chances to be reported, and a pool that is being drawn from is a pool whose bad entries surface.

The consequence is stated on the data deletion page and it is the one to read carefully: deleting your account does not delete the lesson plans you generated. The plan text stays in the pool, with its link to you removed. We could withdraw it instead, and we deliberately do not, because withdrawing it would take a lesson away from someone who is part-way through it. Everything else on that page — your flashcards, your saved words, your progress, your tutor conversations — is deleted with the account.

If that trade is not one you want to make, the thing to know is that it applies to the plan text, which contains no personal data and no link to you once the account is gone. It does not apply to anything you typed into the tutor, which is deleted.

Topics, and what they do

A topic is not a filter. It is a steer. Asking for ordering coffee at a beginner level produces a plan built from the level’s own vocabulary that happens to be set in a café; it does not produce café vocabulary above your level, because the level’s word list is the constraint and the topic is the flavour.

This is why a very narrow topic sometimes produces a plan that seems only loosely connected to it. There may not be twelve words at your level that a plan about filing a tax return could honestly be built from. The generator would rather give you a usable plan that gestures at the topic than an unusable one that insists on it.

Asking for nothing is a perfectly good option, and at low levels it is often the better one: the level’s vocabulary is a curriculum, and letting it choose means covering it rather than orbiting whatever you happened to think of.

When a plan is wrong

Plans are wrong sometimes. A word can be archaic, a sentence stilted, a translation subtly off, a register wrong for the situation. Occasionally something is simply incorrect.

Report it. There is an in-app feedback path on the plan itself, which attaches the plan you were looking at so nobody has to reconstruct it from a description, and there is support@getlingara.com for anything longer. We review every report and act on the ones that need acting on. Because plans are pooled, a report is not just a fix for you — it is a fix for everyone the pool would have served that plan to next.

This matters more than it would in a conventional course, where a mistake is in a book that was printed once. Here the corpus is alive, and reports are the mechanism by which it gets better rather than merely larger.

What the generator will not do

The AI transparency page lists the boundaries in full; two are worth repeating here. Lingara does not certify language proficiency, and it does not make consequential decisions about you. A level in Lingara is a setting that decides which words you are shown. It is not a qualification, it is not an assessment result, and it should not be quoted anywhere that one is wanted.