Teaching Myself Trading with an AI Course System
I wanted to learn trading like a professional, not from YouTube personalities pushing their strategies, but from foundational knowledge drawn from the best traders and theory. No academy I know of offers this, but I knew the knowledge existed, it was just scattered and gated. So I overengineered the problem and built a system that could teach me.
The Architecture
The core idea is clear: create a comprehensive, structured course that could deliver theory, test understanding, and adapt based on performance. I used Claude Code to set up the infrastructure: a folder structure, course theory document, and initial specifications.
The syllabus generation came first. I wrote custom prompts to produce a 180-day curriculum of 15 minute lectures organized in phases: Mechanics, Psychology, Strategy, and Macro knowledge. Each lesson got a title and short description in the syllabus. The length of 15 minutes as well as 180 days was deliberate. Trading has so many layers that you must work from the ground up and build upon each step slowly but surely, and I needed to be able to fit it into my busy fatherhood schedule so I split it up into daily "micro" lessons.
Then I built the specifications for the daily lesson generation itself. Each lesson needed to be TTS-ready. I wanted audio lectures I could consume while multi tasking, or walking with Flynn. After the lesson there would be five questions testing comprehension of the material. A rubric was also generated with specific grading criteria for those specific 5 questions. And finally a teacher prompt that would evaluate answers using that rubric and provide feedback to the user.
Claude Code generated the lessons one by one. The quality held up and as I tested I was surprised by how much I was learning, but I hit token limits fast. After two or three lessons, the system would stop working. Context windows filled up, and token consumption burned through my Claude Pro allocation quickly. I'm now generating lessons as I go rather than building the full 180 days upfront. If I develop this further, I am thinking that just-in-time generation every night for the next day might be more manageable.
The Daily Loop
The delivery system is simple, each morning I get a TTS lecture listen to it carefully and then submit answers through a voice recording that I send to a webhook using a custom tool. The system grades them, provides feedback, and if I've passed, queues up the next lesson for tomorrow.
The grading is the weakest part of the system. The feedback itself is detailed and useful, often pointing to gaps in my reasoning or places where I missed nuance. But the numerical marks cluster around 88/100 regardless of performance. Even when the written feedback is harsh, the score stays in that narrow band. I don't know why, but I have tried a few different things but I think there might be an inherent issue in LLMs where they have a hard time judging how evaluations fit int a scoring system. The rubric is specific, the criteria are clear, but something in the evaluation loop defaults to a good but not perfect score. I have noticed this in other projects too so this is not isolated.
Still, the system works at making me smarter. The lessons have already changed how I think about trading: how I read price action, how I understand risk, what I pay attention to in market structure and how I manage my own psychology.
Why This Matters
Trading sits in a strange category. It's not exactly secret knowledge, but it's gated. No formal education path exists for retail traders. YouTube is full of people selling courses based on their personal approach, which might work for them but often doesn't teach you the underlying mechanics. The incentive structure doesn't favor genuine education.
AI learning systems like this are especially useful when traditional education doesn't serve you: where the knowledge exists but isn't organized into a curriculum, or when you need depth but don't have access to expert instruction.
Meanwhile I'll keep learning, and if this goes somewhere maybe one day I'll make it public.