I have a bad habit with learning. I get interested in something, spend an evening reading about how to learn it, and then never actually learn it. The research feels like progress. It is not.

So when I decided I wanted to get better at something specific this spring — I will keep the exact topic general and just call it "a practical skill I use at work but never formally learned" — I tried something different. I asked AI to build me a learning plan.

Not a reading list. A plan with weeks, sessions, and a way to check whether I was actually improving.

It gave me a beautiful plan. It also gave me a plan that would have failed on day three, and the reason it would have failed is the most useful thing I learned from the whole experiment.

The Task and the Conditions

What I was trying to finish: A realistic 6-week learning plan for a practical skill, with 3–4 sessions per week, each session under 45 minutes, built around free resources.

Tools used:

• ChatGPT (Plus plan) — main plan generation and revision

• Claude (free tier) — second opinion on the revised plan

• A plain notebook — where the plan actually lived, because I did not want to open an app to check my own schedule

Input I gave it:

• My current level: "I can do the basics but I have gaps I cannot name"

• Time available: 3–4 evenings a week, 30–45 minutes each

• Constraint: no paid courses unless clearly worth it

• Goal: be able to do one specific real-world task without looking it up

Time I spent: about 35 minutes across two sessions — one to generate, one to revise after I tried the first version.

The Workflow, Step by Step

Step 1: The first prompt — and the first plan

I asked for a 6-week plan with weekly themes, session-by-session activities, and a way to measure progress.

What came back was genuinely well-organized. Six weeks, each with a theme, each theme broken into 3–4 sessions, each session with a goal and a suggested resource type. It even built in a review week at the end.

If I had stopped there, I would have felt great. That is the problem.

Step 2: I tried week one before trusting the plan

Instead of admiring the plan, I ran the first two sessions exactly as written.

Session one asked me to "review the fundamentals" for 30 minutes. That is not a task. It is a category. I sat there for a few minutes not knowing what to actually open, then defaulted to re-reading something I already knew. I learned nothing and felt behind.

Session two was better — it had a concrete exercise — but it assumed I could already do something the plan had not taught me yet. I spent most of the session stuck on a prerequisite.

So I went back to the tool with real feedback: "Session one was too vague to start. Session two assumed a skill I do not have yet. Here is what actually happened."

Step 3: The revised plan

The second version was noticeably better. Two changes made the difference:

1. Every session got a concrete first action. Not "review fundamentals" but "open [specific type of resource] and write down three things you can already do and one thing you cannot."

2. The prerequisite got moved earlier. The skill session two assumed was now taught in week one, session three.

This version I could actually follow. I completed the first week and a half before I ran out of the schedule I had planned for — which, honestly, is further than most learning plans get me.

What Went Wrong

It built the plan for an imaginary learner

The first plan was written for someone with more time, more background, and more tolerance for ambiguity than I have. It did not know that "review the fundamentals" would paralyze me, because it does not know how I actually behave when I sit down to learn something.

That is not a knowledge gap. It is an imagination gap. The tool can describe a learner. It cannot observe one.

It confused "structured" with "realistic"

The plan looked rigorous. Six weeks, clear themes, a review week. But rigor in the document is not the same as rigor in the schedule. When I mapped the sessions onto my actual week — two evenings already spoken for, one evening I am usually too tired to concentrate — the plan only fit if I ignored my real life.

AI will happily build you a plan for a person with no other commitments. That person does not exist.

It could not tell me when I was actually improving

This is the part it genuinely cannot do for me.

The plan included "checkpoints," but every checkpoint was self-reported. "Do you feel more confident?" "Can you explain this concept?" Those are not measurements. They are vibes.

The real signal that I was improving came from outside the plan: I caught myself doing the task without stopping to look something up. No AI told me that. I noticed it because I was paying attention to my own behavior, not to a checklist.

It kept suggesting resources I had not vetted

The plan pointed me toward articles, videos, and courses. I checked a few. Some were fine. Some were outdated. One was a landing page for a paid product dressed up as a tutorial.

AI does not reliably know whether a resource is current, accurate, or worth your time. It knows the resource exists, or that it sounds like it should. That is a different thing.

What Still Needed My Attention

The honest time accounting: about 20 minutes of AI work, 15 minutes of revision, and then real time spent learning — which the plan could schedule but not do for me.

The Verdict

Use it selectively — for structure, not for motivation or measurement.

I would use AI again to draft a learning plan, because a rough structure beats a blank page. I would not use it to decide whether I am improving, and I would not follow any plan past week one without testing it first.

What it is good at: turning "I want to learn this" into a sequenced set of topics, and suggesting resource types I would not have thought to look for.

What it is not good at: knowing my actual schedule, knowing when a session is too vague to start, and telling the difference between feeling confident and being able to do the thing.

The limitation to remember: A learning plan is a hypothesis about how you will learn. AI can write the hypothesis. Only you can run the experiment. And the first version will probably be wrong — that is normal, not a failure.

The Part It Cannot Do for Me

Here is the honest version.

The plan gave me a map. But the thing that actually moved me forward was sitting down, getting stuck, and noticing where I got stuck. That information is not in the plan. It is not in any prompt. It only exists after you try.

AI can tell you what to learn and in what order. It cannot tell you what you will actually do on a Tuesday night when you are tired and the plan says "review the fundamentals."

That part is mine.

The Anonymized Version of This Workflow

If you want to try this yourself:

"I want to learn [skill] in [timeframe]. I can do [current level] but I have gaps I cannot name. I have [realistic time per week]. I want to be able to [specific real-world task] without looking it up. Build me a session-by-session plan where every session starts with a concrete action, not a category. Flag anything that assumes a prerequisite, and do not suggest paid resources unless there is no free alternative."

Then — and this is the part most people skip — run week one before you trust week two. Report back what actually happened. Revise from there.

Guardrails I Followed

• I did not paste any employer, client, or coworker details into the tool. The skill was described generically.

• I treated every suggested resource as unverified until I checked it myself.

• I did not treat "feeling confident" as evidence of learning.

• I kept the plan in a notebook, not an app, so it stayed in my real life instead of becoming another tab I ignore.

The structure came from AI. The sticking points came from me.

Next in this series: I asked AI to plan a weekend trip. The first itinerary was not ready to leave the house.

I tried it so you don't have to waste your afternoon.