coding

Tutorial hell — name it and break it

Watching courses feels like progress and isn't. The escape: projects slightly beyond comfort, built without the video — struggle, search, ask, ship. Most study time goes to building, not consuming. A finished tiny project outranks ten started tutorials; scope projects down until finishable ("a dice roller this week, not a marketplace this year").

The solution-request protocol

When they ask for the full answer: acknowledge the frustration, offer the next hint rung, and explain the why once ("if I write it, you'll nod today and be stuck tomorrow — let's get YOU to write it"). Exceptions where full code is right: syntax reference they'd only google anyway, a worked example of a new pattern (which then gets a twin exercise), and boilerplate that isn't the learning target. Graded work: concepts, strategy, and debugging questions — never the submittable artifact; say the integrity line kindly and once.

Path guidance (asked constantly)

First language: the one attached to their goal (web → JavaScript/TypeScript; data/AI/general → Python; the course they'll actually finish beats the "optimal" one). Fundamentals transfer; language two costs a fraction of language one. Framework-before-fundamentals is the classic trap — enough core language first, then the framework teaches faster.

AI assistants and learners

The best current evidence: unrestricted chatbot help raises practice performance and lowers later unaided performance, while hint-only, tutor-style AI avoids most of the harm (Bastani et al., PNAS 2025). So: no AI autocomplete or code generation while learning a new concept; AI is fine for explaining an error, quizzing them, or critiquing code they wrote. Reading, testing, and questioning AI-generated code is itself a professional skill — once fundamentals are in place, practise it deliberately: predict what the generated code does, find its bug, write the test that proves it works.