Using Deep Research to Create a Learning System

This post is a note to myself based on this thread by Raphael Traviss.

The main idea is simple:

You can use Deep Research with The Math Academy Way to generate a learning system for almost any skill domain.

Raphael was using this approach to learn traditional art. What stood out to me was not just the use of AI, but the way he described the process. The goal is not to ask an LLM for vague advice like “how do I get better at drawing?” and then receive a generic list of nice-sounding tips.

The goal is to create a system that tells you what to practice, how to practice it, how to measure progress, and how to update the plan once you learn more.

Raphael later shared three example prompts that make the idea clearer:

• Iteration 1
• Iteration 2
• Control prompt

The difference between these prompts is the interesting part.

The Problem With Generic Learning Advice

If you ask an LLM something like:

How do I get better at painting?

You will usually get answers like:

- Practice every day
- Study the fundamentals
- Learn from masters
- Experiment with different styles
- Be patient

This is not wrong, but it is not very useful either.

The advice sounds encouraging, but it does not tell you what to do tomorrow morning when you sit down to practice.

The missing thing is structure.

Use a Learning System as the Prompt

The interesting part from the tweet is the use of Justin Skycak’s article, The Math Academy Way, as the foundation for the prompt.

Instead of asking the model for broad advice, you ask it to design a learning system using specific principles:

• Break the skill into smaller concepts
• Practice the right thing at the right difficulty
• Use repetition over time
• Test understanding frequently
• Review older material before it is forgotten
• Keep adjusting based on weaknesses

This changes the output.

The model is no longer acting like a motivational blog post generator. It has a framework to follow, so it can produce something closer to a curriculum.

Point Deep Research at the Surrounding Material

One detail I missed at first is that the prompt is not only asking the model to read The Math Academy Way.

It also asks Deep Research to look at the surrounding links, discussions, and blog posts that mention it.

That matters because the article itself gives the philosophy, but the surrounding material often explains how the system behaves in practice.

For example, Math Academy is not just a list of topics. It uses a dependency graph of skills and exercises.

That phrase, dependency graph, changes the shape of the answer.

Instead of getting:

Week 1: learn drawing
Week 2: learn values
Week 3: learn color

You want something more like:

Drawing accuracy unlocks value studies.
Value studies unlock limited palette painting.
Limited palette painting unlocks full color portrait studies.
Brushwork and edge control depend on enough paint handling experience.

This is a much better mental model.

It tells you what depends on what.

It also tells you where to go back when something breaks.

If your full color portrait is bad, the problem might not be “color”. It might be that your value structure is weak. If your brushwork is hesitant, the problem might not be the brush. It might be that your drawing is uncertain, so every stroke feels risky.

Example Prompt Shape

The prompt does not need to be magical. The important part is to provide the model with a clear learning philosophy and a target skill.

Something like this:

Read The Math Academy Way and use its principles to create a learning system for learning traditional portrait painting.

Also search for public writing and discussions that explain how the Math Academy system works in practice.

I am a beginner. I want a practice plan that breaks the skill into a dependency graph, gives me exercises, includes review, and tells me how to measure progress.

Do not give generic advice. Give me a system I can follow for the next few weeks.

This kind of prompt gives the model constraints.

It asks for a system, not inspiration.

The Control Prompt Problem

The control prompt is useful because it shows what happens when the Math Academy framing is removed.

The output is still polished. It gives phases, resources, weekly routines, materials, and a reasonable path through portrait painting.

But it mostly behaves like a syllabus.

It says what to study in what order, but it does not make the hidden structure as visible.

The Math Academy-based prompts push the model toward more useful artifacts:

• a hierarchy of sub-skills
• prerequisites between those skills
• concrete exercises for each node
• state changes during the activity itself
• self-assessment checks
• loops for going backward when a prerequisite is weak

That is the real difference.

A syllabus tells you what topic comes next.

A learning system tells you why that topic comes next, what it unlocks, and what failure means.

The First Plan Will Not Be Perfect

Another important point from the thread is that the first version of the learning system is only the beginning.

After following the system for a few weeks, you will start noticing things that were invisible at the beginning.

For example, Raphael mentions discovering that there was a hidden state machine that people were glossing over in the videos he was watching.

That is a very useful observation.

At the start, you might only know:

I am bad at portraits.

After practicing for a while, you might know something more precise:

I struggle to move from construction lines to clean forms, and I do not know when to switch from measuring to simplification.

That second version is much more useful.

Now you can feed that back into the model and ask it to update the learning system.

In the second prompt, Raphael does exactly this. He adds the idea of a state machine and asks the model to pay attention to the internal states of the painting process.

For portrait painting, those states might look like this:

1. Measuring state
  Goal: place the big shapes accurately.
  Trigger to leave: proportions and gesture are correct.

2. Value block-in state
  Goal: separate light and shadow into big masses.
  Trigger to leave: the portrait reads in two or three values.

3. Form modeling state
  Goal: turn flat shapes into believable volume.
  Trigger to leave: the head feels solid and the features are readable.

4. Integration state
  Goal: adjust edges, accents, and overall mood.
  Trigger to finish: nothing pulls attention in the wrong place.

That framing is powerful because many skills are not only about knowing what to do.

They are about knowing which mode you are currently in.

If you are still in measuring state, you should not be polishing eyelashes.

If you are in integration state, you should not keep re-solving the whole drawing unless something structural is wrong.

The Loop

The process looks something like this:

1. Pick a skill
2. Generate a structured learning system using a strong source of learning principles
3. Follow it seriously for a few weeks
4. Notice where you are getting stuck
5. Add those observations back into the prompt
6. Generate the next version of the system
7. Repeat

The important part is step 3.

You have to actually do the reps.

Without practice, there is no feedback. Without feedback, the next prompt is just another guess.

The first prompt is for when you do not know the domain well.

The second prompt is for when you know enough to ask sharper questions.

That is the loop.

Why This Works

I think this approach works because it combines two things:

• The model can organize a large amount of information quickly
• The learner provides real feedback from practice

The AI is useful for creating structure, but the structure gets better only when it is grounded in actual experience.

That is the part I like most about this idea.

It does not treat AI as a shortcut around learning. It treats AI as a way to design better practice.

Summary

The lesson from the thread is:

Do not ask AI for advice.
Ask it to build a learning system.
Then follow the system, collect feedback, and improve the system.

This can probably be applied to programming, writing, drawing, math, music, or anything where progress depends on structured practice.

The key is to put in the time and do the reps.