DEV Community

Cover image for I made my AI tutor stricter: no next lesson until you pass 80%
Valynx
Valynx

Posted on

I made my AI tutor stricter: no next lesson until you pass 80%

#webdev #ai #productivity #learning

published: True

series: Building LearnOptima

Most learning apps remove friction. We added some. In LearnOptima, a product from Valynx, the company I founded, you can't open the next day of your plan until you score 80% on today's quiz.

Here's the reasoning, how it's built, and what I don't know yet.

The problem: a library, not a plan

A video course is a library. Everything you need is in there, and nothing tells you what to do tonight. Anyone who has started one knows the day it stalls: you're tired, the next lesson is long, and "tomorrow" wins.

A plan is different. It says exactly what today is, and it's small enough to finish.

What one small study says

In a preliminary study using a recorded Coursera lecture, people who rewatched it four times did better on a test five minutes later. A week later, the people who watched it once and took recall tests did better. Other studies are mixed, so I treat it as a hint, not a rule. It still shaped the design: every day ends with a quiz, not another rewatch.

How it works

You give LearnOptima a skill, your experience level and how much time you have each day. It generates a 30-day plan for you, fresh each time and not a static template. A 100-day Mastery plan is also available.

Each day has the same shape:

  1. A short text lesson
  2. One hands-on exercise (code, writing or problem-solving)
  3. A quiz

The gate is one rule:

const PASS_MARK = 0.8;

function canUnlockNextDay(correct: number, total: number): boolean {
  return correct / total >= PASS_MARK;
}
Enter fullscreen mode Exit fullscreen mode

That's a simplified version, but the rule really is that simple. No pass, no next day.

Different models for different jobs

LearnOptima routes work through OpenRouter:

Job Model Why
Logic validation and exercise grading GPT-OSS 120B wrong judgments do the most damage here, so it gets the heavier reasoning model
Lesson content and tutoring DeepSeek V4 Flash high volume, so speed and cost matter more

The rule of thumb: spend reasoning where a mistake hurts, and save money where volume is high.

What I don't know yet

  • Whether 80% is the right bar. It's today's setting, not a proven one.
  • Whether short-answer quizzes beat multiple choice here.
  • Whether the gate improves completion at all. LearnOptima is early, so I can't claim that yet.

Your turn

How do you stop LLM graders from drifting on open-ended exercises?

If you want to try it, the first plan is free: learnoptima.online

Top comments (0)