Kite

A personal learning system powered by Claude Code. Open the repo in Claude Code and start talking. Claude builds a knowledge graph of the chosen domain, tracks skill level across topics, quizzes adaptively, recommends what to study next, and generates learning plans and hands-on projects calibrated to current proficiency.

Why I built it

Between my professional work and side projects, a significant portion of my time goes to research: reading, talking with LLMs, exploring ideas to understand them more deeply. The problem is that without recall, most of learning doesn’t persist. A concept feels understood in the moment but becomes difficult to articulate a week later.

Additionally, learning conversations happen across whichever tool is convenient at the time, so notes never accumulate in a single place.

Kite started as a way to force recall and consolidate learning across tools. If I have a knowledge graph of everything I’ve claimed to learn, the system can test me on it and incorporate what I’ve discussed elsewhere.

How it’s built

Kite is interesting in that there’s no traditional code. It’s a collection of markdown instructions, prompt files, and YAML data structures that Claude Code reads and follows. The capability comes from how the prompts are designed, not from a runtime I wrote.

The repo defines:

• A main CLAUDE.md with the workflows Claude follows and the triggers it listens for in natural conversation.
• A prompts/ directory with detailed specs for each workflow: how quizzes are structured, how scoring works, how learning plans are generated, how imports from other LLM conversations get reconciled with the existing graph.
• Three data files Claude maintains: a knowledge graph (topics, relationships, priorities, difficulty), a skills file (proficiency level on each topic from 0 to 4), and an append-only learning log.
• A paired extract/import flow. At the end of a learning conversation in any LLM (ChatGPT, Gemini, Claude.ai, anything), paste the extract prompt. The model outputs a structured YAML summary of what was discussed. Paste the YAML back into Kite, and the import prompt tells Claude how to reconcile new information against the existing knowledge graph: match new topics to existing topics, update skills, log changes. Learning stays consolidated regardless of where conversations happen.

On first use, Claude detects the uninitialized state and runs a five-minute interview to set up the domain. After setup, Claude responds to natural triggers: “status,” “quiz me on X,” “what should I learn next,” “learning plan for X,” “import.”

Why it’s a public template

Kite is a template meant to be cloned into a personal instance, not a hosted service. Everything is in plain text, so anyone can fork the repo, configure it for a different domain, and customize the prompts to change how quizzes, scoring, or plans work. I wanted others to be able to use Kite without me running infrastructure.

Stack

Claude Code. Markdown for instructions. YAML for data. No dependencies, no scripts.