How to build a personal curriculum with AI that actually sticks

Ask any chatbot for a "30-day plan to learn X" and you get something impressive in seconds: weeks, topics, a few resources per day, maybe a motivational closing line. It looks like a curriculum. You feel organized. You might even start.

Then it rots. By day four the plan is a tab you do not open, the early topics have already faded, and you cannot tell whether you are behind or ahead because nothing ever checked. This is the quiet failure mode of self-directed learning with AI: generating a plan has become free, and finishing one has not.

If you searched for how to build a personal curriculum with AI, you have probably already made three of these plans. The problem is not the AI. It is that a list of topics was never the hard part.

A plan is not a curriculum

A real curriculum, the kind a good course or a good teacher gives you, is not a list. It is a loop with four things a topic list does not have:

• A sequence. Ideas in an order where each one earns the next, not a flat pile of subjects.
• Retrieval. Moments where you are asked to produce what you learned from memory, not just reread it.
• Feedback. Something that tells you what you actually got wrong, not your own guess about it.
• Spacing. A reason to come back to last week's material before it evaporates, instead of marching forward and leaving a trail of forgotten units.

A chatbot plan gives you the first one, badly, and none of the other three. That is why it feels productive and leaves nothing behind. We wrote about the underlying trap in why notes and reading do not stick: recognizing an answer feels like knowing it, and a plan made of "read this, watch that" is recognition all the way down.

Where AI actually helps, and where it quietly fools you

AI is genuinely transformative for one half of this and useless for the other, and it is worth being honest about which is which.

Where it helps: mapping the terrain. AI is excellent at taking a fuzzy goal and turning it into a sane sequence, naming the prerequisites you did not know you were missing, suggesting what to skip, and explaining any single concept at exactly the depth you ask for. The thing that used to require a syllabus, a textbook's table of contents, or a mentor is now a five-minute conversation.

Where it fools you: the loop. By default, AI does the remembering for you. You ask, it answers, the answer is on screen, and your brain files it as known. There is no moment where it hides the material and makes you produce it, no honest grade, no schedule that drags week one back into view during week three. Left alone, an AI tutor is the most convincing version of the fluency illusion ever built, because it is so good at handing you the answer the instant you feel a gap.

So the move is not "ask AI to teach me." It is: use AI for the map, and build the loop on purpose.

How to build it, step by step

1. Define the outcome, not the topic

Start from what you want to be able to do, not what you want to "learn about." "Understand statistics" rots. "Be able to read a study and say whether its conclusion is justified" gives the whole curriculum a target to aim at, and gives you a way to know when you are done. Tell the AI the capability, not the subject, and the plan it returns gets sharper immediately.

2. Let AI draft the sequence, then prune it hard

Ask for the dependency order: what must come before what, and what is optional. Then cut. The first plan an AI gives you is almost always too long and too even, every topic weighted the same. A real curriculum is lumpy. Three ideas usually carry most of the subject. Ask which three, and let the rest be supporting cast.

3. Turn each unit into a page you build, not a link you collect

This is the step people skip, and it is the one that matters. Instead of bookmarking a video and an article per unit, create a page for the unit and put the actual material into it: write the explanation in your own words, or import the source you are learning from. The act of assembling the unit is the first pass of real learning, and it gives you something concrete to be tested on later. A folder of links is not a curriculum. A set of pages you built is.

4. Generate recall and practice from the material, not more reading

Once a unit exists as a page, turn it into the thing AI is bad at giving you by default: questions you have to answer from memory. Good recall is not "define X." It is "explain why X, in your own words," then something that checks your answer against the source and tells you what you missed. For skills, you want a drill or a short back-and-forth, not a definition. This is the same wedge we drew comparing it to flashcard apps: the cards are generated for you, but the recall is graded, not just generated.

5. Put the whole thing on a spaced schedule

This is the difference between a curriculum and a pile of finished units. Each unit should come back, automatically, right before you would forget it, with the intervals stretching as you prove you remember. That is what spaced repetition with a modern algorithm like FSRS does, and it is the mechanism that turns "I covered that in week one" into "I still have that in month three." Marching forward without it means every unit you finish starts decaying the moment you move on.

6. Treat the curriculum as living

You will be wrong about the sequence. A unit you thought was foundational turns out to be a detour; a topic you skipped turns out to be load-bearing. The advantage of building your own curriculum with AI is that editing it costs nothing. Re-sequence when reality disagrees with your plan, split a unit that is too big, drop one that is not earning its place. A static 30-day plan cannot do this. A living one should.

The trap: a study plan from a chatbot

To be concrete about what to avoid, here is the default path and why it fails:

The chatbot plan is not useless. It is a great first draft of column one. The point is to not stop there.

A worked example

Say you want to learn enough machine learning to evaluate a model someone hands you, not to publish papers.

• Outcome: "Given a model and its metrics, I can say whether it is trustworthy and where it will fail."
• Sequence (AI drafts, you prune): what a model is actually optimizing, the train and test split and why it matters, overfitting, the handful of metrics that matter for your case, and how they lie. You cut the deep math and the history.
• Pages: one per unit, written or imported from the one course you picked, not five tabs.
• Recall: not "define overfitting," but "here is a model that scores 0.99 on training and 0.6 in production. Explain what happened." Answer from memory, get graded.
• Schedule: the overfitting unit resurfaces a few days later, then a week, then a month, while you are working on metrics.
• Living: you discover you needed a unit on data leakage. Add it where it belongs and keep going.

Six weeks later you do not have a finished plan you forgot. You have a capability that still answers when the material is closed.

Where this fits

You can run this loop with a stack of separate tools: a chatbot for the map, a document for each unit, a flashcard app for recall, and a calendar for spacing. It works, and the friction of stitching them together is exactly why most people quit by day four.

Anti-Agent is built to be the single surface for it. You write or import each unit as a page, the page turns the material into graded recall and practice instead of more reading, and everything lands on an FSRS schedule that brings it back on its own. It is the same idea as turning sources into spaced recall instead of a one-time chat, applied to a whole self-built curriculum rather than a single document.

Frequently asked questions

Can AI build a personalized learning curriculum for me? It can build the map: a sane sequence, the prerequisites, the depth you ask for, in seconds. What it does not do by default is the part that makes a curriculum work, which is testing your recall, grading it honestly, and bringing material back on a schedule. Build the plan with AI, but build the retrieval and spacing loop on purpose around it.

Why do the study plans ChatGPT gives me never work? Because a list of topics and links is the easy 10 percent. It has a sequence but no retrieval, no feedback, and no spacing, so the early units fade while you march forward and nothing ever checks whether you actually learned anything. The plan is fine. The missing loop is the problem.

How is this different from just asking an AI tutor questions? An AI tutor answers in the moment, which feels like learning and usually is not, because the answer is always on screen. A curriculum hides the material and makes you produce it from memory, grades that, and schedules the next attempt. Conversation is part of it, but it is not a substitute for retrieval over time.

Do I have to design the whole curriculum up front? No, and you should not. Define the outcome, draft a rough sequence, and start building the first few units. Treat it as living and re-sequence as you learn what actually matters. A curriculum you adjust beats a perfect plan you abandon.

Is there a tool that does the recall and scheduling part for me? Yes. Anti-Agent generates recall and practice from each unit you build, grades your answers against the material, and schedules everything with FSRS, so the only thing you own is deciding what to learn and assembling the units. There is a free tier to run the full loop.

The bottom line

AI made the cheap part of learning, generating a plan, effectively free. That is a real gift, and it is also a trap, because a plan you can produce in ten seconds is a plan you will abandon in four days when it turns out a topic list was never going to teach you anything.

A personal curriculum that sticks is a loop, not a list: an outcome, a pruned sequence, units you build, recall that is graded, a schedule that brings it all back, and the freedom to re-sequence when you were wrong. Use AI for the map. Build the loop on purpose.

Start your first unit as a page and let it come back to you on a schedule.