Frequently Asked Questions About Jason Ku AI-Native Learn-to-Code Framework
21 answers covering everything from basics to advanced usage.
// Basics
What exactly is a Knowledge Graph in this framework?
The Knowledge Graph is a living, agent-maintained record of every concept you've encountered, each tagged with a mastery status — Introduced, Practicing, Understood, or Mastered — plus an evidence log of completed tasks. The agent reads it before each task to calibrate how deeply to explain or quiz you, so it never over-explains basics or skips essential foundations.
What's the difference between a coding agent and a vibe-coding tool?
A coding agent (Claude Code, Codex, Cursor) writes, runs, and iterates on code interactively while exposing the underlying code and concepts to you. Vibe-coding tools (Lovable, Base44, Replit) abstract away that complexity to maximize ease of use. The framework insists on agents because the skills they build transfer to real engineering jobs, while vibe-coding skills don't.
What does 'Learning Happens in the Friction' mean?
It's the core pedagogical principle that real understanding comes from deliberate slowdown — predicting code behavior before running it, answering quiz questions, and reading every line — not from letting AI generate solutions unexamined. Speed is the enemy at the learning stage. The uncomfortable effort of thinking through each line is precisely where knowledge forms.
What is the trunk of the Knowledge Tree?
The trunk is the core system components common to almost every software project: source control (Git), backend, database, frontend, API, and deployment. You define it before writing any code, so you get a high-level map of what you're building. Branches and leaves attach to this structural scaffold, which is what makes individual concepts memorable.
// How To
How do I set up my first coding agent?
Pick one of Claude Code, Codex, or Cursor (Claude Code for beginners), skip free tiers and vibe-coding platforms, and get the $20 plan. Spend one evening installing it and watching one or two tutorials covering context management, memory, model selection, and plan mode. Don't try to learn everything — just enough to start your project.
How do I break my project into branches?
Have the agent split your MVP into horizontal feature slices that build sequentially on each other, each being a shippable increment. A typical sequence: (1) serve a plain HTML page, (2) add interactivity with JavaScript or React, (3) add persistence with a database, (4) deploy. These become the branches of your Knowledge Tree.
How do I run a single task so I actually learn from it?
For each leaf task: the agent explains what the task does, generates code, then asks you to describe in your own words what you expect to happen before running it. Run it and review together, then let the agent quiz you on every new concept. Afterward, update the Knowledge Graph with mastery evidence.
How do I initialize the Knowledge Graph?
In Step 6, as the agent defines each task it must also identify every underlying concept the task introduces — variables, arrays, Git commits, API endpoints. For each, create an entry with the concept name, a starting mastery status of 'Introduced,' and an evidence list. This graph then guides explanation depth and quiz intensity for the rest of the project.
// Troubleshooting
What do I do when I'm tired and tempted to skip the quizzes?
Treat the temptation as a signal to stop the session, not to lower the bar. Skipping quizzes or prompting the agent to 'just fix it' is the precise habit that prevents real skill development. End the session, rest, and resume when you can engage properly. Never trade understanding for speed at this stage.
My project stalled before reaching the MVP — what went wrong?
You likely picked a first project that was too complex — a two-sided marketplace, payment flow, or multi-role system. These stall before MVP and kill motivation. Strip your project down with the agent to the smallest end-to-end working version, push everything else to the backlog, and ship the core loop first.
The AI keeps generating code I don't understand — how do I fix that?
Enforce the predict-before-run rule: never let the agent run code until you've stated what you expect to happen. Require it to quiz you on each new concept and refuse 'just fix it' prompting. If your codebase is already spiraling, slow down, rebuild the Knowledge Graph for what's there, and walk back through each layer with the agent.
I keep forgetting concepts I learned last week — what should I change?
You're probably learning loose, unanchored leaves. Make sure every concept is attached to a branch and trunk inside your project, and that the Knowledge Graph logs evidence of review. Durable knowledge comes from relationships between concepts within a system you're building, not from isolated facts memorized in a tutorial.
// Comparisons
How does this compare to learning a language in isolation first?
Learning JavaScript or Python 'first' before having a project creates loose, unanchored leaves that are quickly forgotten. This framework flips that: you pick a project first, then learn each language concept exactly when a task needs it. Concepts are immediately contextualized and reinforced by real use, making them far stickier.
How does this compare to just vibe-coding an app with Replit or Lovable?
Vibe-coding tools get you a working app faster but hide the complexity you need to understand, so you can't explain your own code and the skills don't transfer to a real job. This framework is slower by design but produces genuine engineering ability — you finish able to explain every architectural layer of a deployed app.
How does this compare to traditional computer science degrees?
A CS degree is theory-heavy and front-loads abstraction before application. This framework is project-first and AI-native: you develop coding fundamentals and AI collaboration skills in parallel from day one by shipping real software. It won't cover deep theory like algorithms at depth, but it gets a beginner to a deployed, explainable app far faster.
Is Claude Code or Codex better for me?
Claude Code is better if you're a true beginner — it's the most interactive, which supports the quiz-and-predict loop central to this method. Codex is more autonomous and suits experienced users who want the agent to do more independently. Cursor is the third professional option. All three are acceptable; match the choice to your experience level.
// Advanced
How does the agent decide how deeply to explain a concept?
It consults the Knowledge Graph before each task. If a concept like 'variable' is still 'Introduced,' it explains fully and quizzes gently. By the time that concept is 'Practicing,' it skips the basic explanation and goes straight to a harder quiz. This adaptive calibration prevents both over-explaining basics and skipping foundations.
What is a Coding Harness and do I need to build one?
The Coding Harness is Jason Ku's term for the broader system wrapping the coding agent — the interface, step tracker, Knowledge Graph visualization, and session management that keeps you oriented. You don't have to build one from scratch; you can approximate it with structured notes and consistent agent instructions, but a dedicated harness makes the method smoother.
How do I calibrate my timeline based on available weekly hours?
Share your rough weekly hours with the agent when scoping the project and MVP so it can set realistic expectations and size branches accordingly. Fewer hours means a tighter MVP and smaller leaf tasks per session. The goal is still to ship the core loop in weeks, so scope the MVP down rather than extending indefinitely.
Can I use this framework to restart after quitting coding before?
Yes — it's explicitly designed for restarts. Previous attempts usually fail because they were collections of dangling leaves with no tree. Starting fresh with a project you genuinely care about attaches every concept to a system, and the MVP scoping closes the motivational feedback loop that tutorial-hopping never provided, so this time it feels real.
How do I know when a concept is truly 'Mastered'?
A concept moves to 'Mastered' when the Knowledge Graph shows repeated evidence across multiple tasks and you can answer harder quiz questions about it without explanation. The progression is Introduced → Practicing → Understood → Mastered, driven by accumulated evidence of correct prediction and application, not by a single correct answer.