Proxima Learn

Inspiration

I first built this idea as a small web based AI tutor to help me learn faster. It worked, but I noticed that I still had to move between the tutor, my notes, my code, and the actual project I was working on.

I kept changing the tutor based on how I used it. Eventually, I realised that a better place for it was inside VS Code, where most of my learning and building already happens. That earlier web app gradually evolved into Proxima Learn.

What it does

Proxima Learn is a VS Code extension that uses Codex as a learning partner.

It does not always give the full answer immediately. It can ask me to try first, give a small hint, check my attempt, and provide more help when I need it.

It works with files, code, notes, screenshots, and other material already in the workspace. It also keeps a learning state file with the current goal, progress, difficulties, and things that need review. This helps the learning continue across multiple conversations instead of starting from zero every time.

It has three modes. Explore is for understanding a topic through curiosity. Coach is for serious practice and improvement. Agent is for situations where I simply need something completed.

How I built it

I built Proxima Learn as a TypeScript VS Code extension with a React interface.

Codex handles the model conversations, tools, file access, approvals, and commands. Proxima adds the learning prompt, methodology, persistent learning state, and the interface around those capabilities.

I used Codex throughout development to understand the architecture, write and debug code, create tests, improve the interface, and package the extension. I tested the learning behaviour mainly with GPT 5.6 Sol.

The first working version came together in about one intensive day. After that, I spent more time fixing the UI, testing real conversations, improving the prompt, handling Windows and macOS differences, and making the extension easier to install and use.

Challenges I ran into

Initially, I thought I would need to create a VS Code fork and build an entire AI assistant interface from the beginning.

Later, I found that I could use the existing Codex runtime and focus on the part that was different about Proxima Learn. This saved a lot of time and let me work more on the actual learning experience.

The interface was difficult for me because I do not have much UI design experience. Streaming messages, showing commands, displaying reasoning activity, handling approvals, showing file changes, and making everything work at different window sizes took several attempts.

Supporting different machines was another challenge. Something that worked on macOS did not always work the same way on Windows. Finding the Codex runtime, handling paths, resuming conversations, and packaging the extension all needed additional testing.

Accomplishments that I'm proud of

I am happy that the tutor finally found the right form. It is no longer just a prompt or a separate chat experiment. It is now an installable extension that works inside a real workspace.

It has persistent learning state, learning modes, file and image attachments, conversation history, approvals, model controls, rendered mathematics and diagrams, and a UI built around learning activity.

I also did not expect Codex to make development this fast. A functional extension was possible in about one day, although making it reliable and presentable took longer. The experience showed me how quickly an idea can become something usable when the development tool can help with architecture, implementation, debugging, and testing.

What I learned

I learned that a useful AI tutor needs more than a good prompt.

It needs memory, a clear learning method, access to the learner's real work, and an interface that makes its actions understandable. The small details around permissions, file changes, history, and feedback matter as much as the model response.

I also learned that using an existing runtime is much better than rebuilding everything. Codex already provided the model and tool infrastructure, so I could spend more time on the learning behaviour.

What's next for Proxima Learn

I plan to use Proxima Learn regularly for my own learning and see where it actually helps and where it becomes frustrating.

I will continue changing the prompt and methodology based on real sessions. I also want other people to try it and tell me whether it helps them understand and remember more.

I will add new features when regular use shows that they are needed. I also plan to keep improving platform support, reliability, and the installation experience.

Built With

• claude
• codex
• vscode