I'm Elena, a soft skills instructional designer for corporate clients at an online school. In four years in EdTech, I've built 27 courses from scratch — from the first draft structure to final LMS export. Before, an 8-lesson course took me 40 hours of pure work and 2-3 calendar weeks. Now, I make the same course in 8 hours and 4 workdays — with the same quality, and a better presentation.
It's not magic, and it's not cutting corners. It's a rebuilt workflow around AI in one Telegram bot. Below is a step-by-step breakdown of my process: which models work at what stages, what I generate fully with AI, where I still do things manually, and how many credits a typical 8-lesson course uses.
Step 1. Lecture Content from a Bullet Plan
The hardest part of any course is writing eight lecture texts, each 1500-2000 words long. That used to take me 16 of those 40 hours. Now, **Claude 4.5 Sonnet** in Quantium handles 12 of those hours. It holds long contexts and narrative logic for educational materials better than GPT-5.
The algorithm: the expert author and I sit for an hour, building a bullet plan for each lecture — 12-15 points with examples, facts, and intros. It's not text; it's a skeleton. Then, the prompt for Claude:
Claude gives me a draft in 90 seconds. The expert and I spend 30 minutes reviewing it and refining the facts. That's 35-40 minutes per lesson instead of two hours. For an 8-lesson course, it's 5 hours instead of 16.
Step 2. Lesson Covers in a Series
Covers used to be drawn by an in-house designer: 1.5 hours for 8 covers in a consistent style, costing 5000-8000₽ per course based on my internal billing. Now, I make them myself in 50 minutes using **Gemini Image**.
My method: First, I generate one reference cover with a detailed style description. Once I like it, I copy the prompt template and just change the subject. For a consistent series, a template like this works:
I only change the part in brackets: 'two people talking at a table,' 'manager and team in a meeting,' 'person looking at a results graph.' Fifty covers in a consistent style in an hour is totally doable. Learn more about working with Gemini Image in our Midjourney alternatives comparison.
Step 3. Tests with Explanations for Wrong Answers
A good test isn't just 'pick the right answer,' it's 'pick the right answer and understand why the others are wrong.' Before, I'd write these tests for an hour per lesson: question + 4 answers + an explanation for each. 8 lessons × 1 hour = 8 hours of work.
Now, **Claude 4.5** handles this block entirely. Here's the prompt:
I get a ready test in 60 seconds, then spend 10 minutes checking and editing. 8 lessons × 15 minutes = 2 hours instead of 8. And the test quality is often better than mine — Claude catches typical cognitive biases I'd miss.
Step 4. Script Voiceovers via TTS
Video lessons need voiceovers. Before, I'd either hire a voice actor (2000₽/min of finished audio) or record it myself and then clean up the noise. For 8 lessons at 5 minutes each, that's 40 minutes of audio, 80,000₽, or 8 hours of my time.
Now
"Students will be mad the course was made by AI." They weren't. The course isn't "made" by AI — it's made by an instructional designer using AI. It's like saying "the doctor treated them with an MRI." Nobody thinks the MRI does the treating.
"The tests will be superficial." Turns out, the opposite: Claude generates tests on common cognitive biases better than most instructional designers. The key is giving the model the full lecture text in context, not just a description.
"This will save hours but eat up the AI budget." The actual cost for 10-15 courses a year with a Premium subscription is less than a junior instructional designer's monthly salary. So we didn't "add an expense"; we replaced part of the team that simply couldn't keep up.
Learn more about our plans on the pricing page. Related materials: marketer's case study, FLUX prompt guide.
Frequently Asked Questions
Can AI build a full course without an expert?
No. AI assembles the structure, lecture material, and tests from your bullet-point plan, but you still need an expert. They verify facts, add real-world case studies, and adapt content for practical use. Without an expert, the course will feel encyclopedic and lack depth.
What tasks in an online school can AI fully handle?
Lesson covers, tests with answer explanations, checklists, preview videos for ads, webinar transcripts, basic lecture texts from a bullet-point plan, subtitle translation. That's about 60-70% of an instructional designer's work that used to be done manually.
What models do instructional designers use in Quantium?
Claude 4.5 for lecture texts and tests (it handles long context and logic better), GPT-5 for copywriting and descriptions, Gemini Image and FLUX 2 Pro for lesson covers, ElevenLabs TTS for voiceovers, Sora 2 and Veo 3 for ad teasers. Everything's available with one subscription in the Telegram bot.
How many hours does AI really save per course?
Based on our case study, it's from 40 hours per course down to 8-10 hours. That 30-hour saving comes from three areas: lecture material (16 down to 5 hours), covers and visuals (3 down to 1 hour), and tests with explanations (8 down to 2 hours). Expert review still takes 2-3 hours.
How do you generate a series of lesson covers in a consistent style?
In Gemini Image, you give it one cover as a reference and ask it to generate the rest in the same style, only changing the text content. This also works in FLUX 2 Pro using a strict prompt template with fixed parameters for color, composition, and style. You can realistically make 50 covers in a series in an hour.
Try Quantium for Free
20 credits per month on the free plan. 30+ AI models in one Telegram bot.
Open bot →Read also
