Demo videos
OpenAI Build Week — Codex implementation demo
Watch the Codex demo on YouTube
Shows the Codex-built repository workflow, learner-owned control layer, and how the project evolved for OpenAI Build Week.
AceSAT Education AI-Agent — Digital SAT Math demo
Watch the Digital SAT Math agent demo
Shows Bluebook Math Module 1 feedback import, percentage-misconception diagnosis, equation and reasoning writeback, support tracking, and delayed transfer review.
Inspiration
GitLearn OS grew out of my own exam preparation. My learning was spread across textbooks, class notes, teacher feedback, Bluebook results, practice mistakes, unfinished tasks, and separate AI conversations. Each tool could help for a moment, but the learning process itself did not remain connected.
I first used a GitHub repository to organize learning materials. That solved storage, but not coordination. A repository could hold files, yet it could not interpret new evidence, update the learner's state, retire obsolete work, or explain why the next plan changed.
What it does
GitLearn OS is a learner-owned control layer for long-term learning. One replaceable main AI agent coordinates events from school, self-study, teachers, textbooks, assessment platforms, and AI tools, while durable state remains in the learner's own GitHub repository.
The system can:
- import notes, mistakes, teacher feedback, and assessment results;
- connect new evidence to goals and knowledge gaps;
- generate the smallest useful diagnostic or targeted practice item;
- save the learner's answer, reasoning, and support level instead of only a score;
- update mastery and next actions from evidence rather than confidence alone;
- schedule variation, transfer, and delayed-review checks;
- produce inspectable, reversible Git writeback receipts.
A typical repository keeps shared policy and cross-subject planning at the root, while subject-specific state lives under subjects/<subject>/. Git history records not only the current learning state, but how and why it changed.
Digital SAT Math demo
The AceSAT demo starts with real-style Bluebook Math Module 1 feedback:
I missed two percent problems. I multiplied the final value by 0.8 both times.
GitLearn OS does not respond with a long lesson or a generic worksheet. It identifies one likely misconception: the learner has not consistently identified the base quantity before applying a percentage multiplier. It then creates one fresh check:
After a 20% discount, a jacket costs $72. What was the original price?
The learner explains that 0.8 multiplies the original price, writes 0.8p = 72, and solves p = 90 without a hint. The system writes back the equation, reasoning, support used, and the rule "identify the base quantity before applying a multiplier."
One immediate correct answer is not treated as mastery. The state remains learning, and GitLearn OS schedules a 48-hour transfer check in a new context using the same model and no hint. The delayed transfer result becomes the stronger evidence.
How it works
For each learning event, the agent follows an evidence loop:
- Read the existing learner-owned state.
- Interpret the new input as evidence, not as a complete truth.
- Choose the smallest useful next action.
- Capture the learner's response, reasoning, and support level.
- Write the justified update and next check into Git.
- Return a concise receipt of what changed and why.
A tool-capable agent can perform safe, reversible organization, question generation, synchronization, and writeback. A read-only agent must instead return exact pending writeback and clearly state that no repository change occurred.
How I used AI tools
ChatGPT helped reshape the project from a collection of learning files into a control layer that separates conversation, reasoning, persistent learning state, and Git history.
Codex helped turn those product decisions into the repository, documentation, skills, templates, website, and repeated implementation refinements.
GPT-5.6 provides the reasoning inside the learning workflow: interpreting intent, connecting evidence to existing state, identifying gaps, selecting the smallest useful next action, generating targeted checks, and explaining why a state update is justified.
The main agent is replaceable. The workflow can run through ChatGPT Work, Codex, OpenCode, or another tool-capable AI; the durable learner state and its history stay in Git.
Why it is different
For a quick answer, a normal chatbot may be enough. GitLearn OS becomes valuable when learning continues for weeks or months and involves multiple assessments, people, materials, goals, and changing evidence.
It does not replace teachers, textbooks, classes, Bluebook, or AI applications. It helps those parts work together while keeping the learner in control of the durable state.
Try it
Read QUICKSTART.md, create or choose a private target repository, connect it to a tool-capable main agent, and provide one subject, learning goal, and current learning input. The agent should preserve existing content, create only the state needed for the current event, perform safe reversible work, and finish with a concise receipt of changes, evidence, automation, and the next action.
