🚫 Never ghostwrites — a hard constraint. The agent will not complete, fill in, or hand over a single line of your code. Mid-exercise help explains the concept with a worked example in a different domain and different variable names, so the answer can't leak. The real solution appears only in the post-submission review — and trying to fish for it by leaving learn mode doesn't work either.

🎓 A four-beat teaching loop. Every lesson runs Tell → Show → You Do → I Check: teach from first principles (jargon defined before use), show a runnable, commented example in a different scenario than the exercise, hand you a blank file to write yourself, then run your code and grade the real output against the requirement — because "it ran" is never the same as "it's correct."

📊 Evidence-based mastery tracking. Every skill is scored on a 0–100 rubric (Correctness · Tests · Standards · Readability · Concept grasp) and moved through not started → learning → practiced → mastered only on objective evidence: a skill is mastered only after two clean, un-hinted reps on separate, spaced occasions. No self-certification.

🚦 Advancement gates that won't let you skip. You move past a lesson only when you score ≥ 60 and have no blocking skills still below practiced. Weak areas trigger targeted re-teaching in a fresh scenario; proven strength triggers acceleration — compress drills, offer a test-out. The pace adapts in both directions.

🔬 Reads the docs with you, not for you. Every library lesson teaches you to decode a real VS Code hover signature into plain English yourself — token by token, rating your own confidence — so your reliance on AI drops instead of growing.

📡 No more outdated tutorials. No more deprecated APIs. Before any library lesson, the current API is pulled live from the official docs via Context7 (MCP) and pinned to the non-deprecated version — package names verified too, not just methods. You learn the framework as it exists today, not as some three-year-old blog post remembers it.

🎚️ Don't like how a lesson lands? Just say so. Tell the agent how you learn best — visual-first, terse, more examples, slower steps, a different analogy — and it adapts its style, depth, and pace on the fly. The lesson bends to you, not the other way around.

📈 Spaced-repetition + adaptive drills. Standing commands — drill me, leet me, test your skills — pull what's due from a spacing schedule and weight toward your current weak spots. A read-only how am I doing? gives an honest snapshot of scores, gaps, and what's next, any time.

🛡️ Learner guardrails that watch for trouble. Silent detectors for frustration, copy-paste / AI-submitted work (met with "walk me through line N"), perfectionism (every task is time-boxed), and vocabulary gaps — each one changes how the agent teaches on the fly.

💼 Taught to what employers are actually hiring for. A market check command pulls live signal from the Stack Overflow Survey, HN "Who is Hiring," roadmap.sh, and LinkedIn job listings, then maps your skills against what AI-app roles are actually hiring for — and weights the curriculum's depth toward high-demand, within-stack skills. You learn where the job market spends its time, not what's merely fashionable.

✋ Git stays in your hands. You run every git command yourself (git init, add, commit, switch, merge); the agent teaches and verifies but never commits for you. Every submission is your own commit on a feature branch.

🗺️ A 16-module curriculum to employment-ready. A fixed dependency ladder from JavaScript and Git fundamentals through TypeScript, React, full-stack Next.js, databases, auth, testing, security, and CI/CD — ending in a three-app AI portfolio (LLM apps, RAG, and agents) and interview readiness in the AI niche. You ship your first live page in week one.