I had Claude build me 12 full courses. Here is the skill that does it
In 2026, it is hard to hold our attention and learn anything deeply. We ask AI agents our questions and forget the answers by morning. We rarely have the quality time ittakes to read a book and understand a topic deeply.
There are no learning paths personalised and curated for us, and no deliberate practice that proves you actually know something instead of only being able to look it up on Claude.
So I built Course Creator, a skill for Claude Code and Codex that turns anything you want to learn into a real course. https://github.com/ShivamGupta42/course-creator
You give it a topic. It builds a full course you open in your browser: a path from beginner to advanced, every idea rebuilt from first principles, a hands-on project in every lesson, and quizzes that explain why you were wrong.
I have already used it to build 12. You can click into any one right now, nothing to install:
Live: https://shivamgupta42.github.io/course-library/
What it actually gives you:
- A clear learning path. 25 lessons in three tracks, beginner to advanced, instead of a single answer that leaves you where you started.
- You understand, you do not memorize. Every idea is rebuilt from first principles, Feynman style, then explained back in five plain sentences.
- You build something every lesson. A hands-on project ends each lesson, plus interactive labs you can play with.
- It explains the reasoning behind common mistakes. Get a quiz wrong and it tells you which idea you confused and why the right answer wins, instead of just "incorrect."
- You grade yourself. Every project comes with a rubric and a pass line, so you know when you have actually got it.
- It is tuned to you. Tell it who it is for and how deep to go, and it builds for that.
For e.g. uses the following image that is part of system design course to you if you are a beginner
What is the one thing you have always meant to learn and never got to?
Both repos are open source:
- The skill: https://github.com/ShivamGupta42/course-creator
- The 12 courses: https://github.com/ShivamGupta42/course-library
PRs and ideas are welcome :)
I built a plugin that does better loop engineering than claude
𝐀 𝐥𝐨𝐨𝐩 𝐧𝐞𝐞𝐝𝐬 𝐦𝐨𝐫𝐞 𝐭𝐡𝐚𝐧 𝐚 𝐩𝐫𝐨𝐦𝐩𝐭.
It needs to gather the right context, solve the actual problem, and build things the way you like.
It needs a plan it can carry across sessions, a way to check the work against your real goal, and a record of every decision it made while you were away.
That is what I am building with 𝐆𝐨𝐚𝐥𝐬. https://github.com/ShivamGupta42/goals
𝐆𝐨𝐚𝐥𝐬 is a small CLI + plugin for Claude Code and Codex. Your agent gets commands to create the goal, take the next step, check status, view a dashboard, import a proven loop, and improve the loop itself.
Everything it does is saved in 𝐟𝐢𝐥𝐞𝐬 𝐲𝐨𝐮 𝐨𝐰𝐧: the goal, the current phase, the decisions, the evidence, the failed checks, and the history.
Can we make agents run longer? We can.
The better question: can a long-running agent build something you can 𝐭𝐫𝐮𝐬𝐭, 𝐯𝐞𝐫𝐢𝐟𝐲, 𝐟𝐢𝐱, 𝐫𝐞𝐮𝐬𝐞, 𝐚𝐧𝐝 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐞?
That is where Goals comes in:
- 𝐏𝐥𝐚𝐢𝐧-𝐄𝐧𝐠𝐥𝐢𝐬𝐡 𝐠𝐨𝐚𝐥𝐬: say what you want and Goals turns it into tracked phases.
- 𝐕𝐞𝐫𝐢𝐟𝐢𝐞𝐝 𝐬𝐭𝐞𝐩𝐬: a step is accepted only after its proof actually runs.
- 𝐃𝐞𝐜𝐢𝐬𝐢𝐨𝐧𝐬 𝐲𝐨𝐮 𝐜𝐚𝐧 𝐫𝐞𝐚𝐝: every choice is explained by goal, risk, and reversibility, and you can change it.
- 𝐅𝐢𝐱𝐚𝐛𝐥𝐞 𝐟𝐚𝐢𝐥𝐮𝐫𝐞𝐬: a failed check points to the next repair instead of a vague retry.
- 𝐑𝐞𝐮𝐬𝐚𝐛𝐥𝐞 𝐥𝐨𝐨𝐩𝐬: import a proven loop, fill in the gaps, and Goals validates it before you run it.
- 𝐑𝐞𝐬𝐮𝐦𝐚𝐛𝐥𝐞 𝐰𝐨𝐫𝐤: pick up after a /clear, a new session, or a switch between Claude Code and Codex.
The project is open source:
https://github.com/ShivamGupta42/goals
If you are running long loops, try Goals on one real project and tell me where it falls short. PRs and suggestions are welcome :)