AI as a coding tutor, not just a code generator

A well-engaged HN thread pushed back on the dominant narrative of AI as a replacement for thinking. The author described using LLMs not to generate finished code but as a tireless tutor: writing intentionally rough code, then working through the corrections in a long back-and-forth. Claude Code's /code-review command got a specific callout as useful. The framing is deliberately slow and deliberate, which is the opposite of how AI coding tools are usually marketed.

This sits in productive tension with another HN thread lamenting that nobody reads programming books anymore, which argued that junior developers who skip foundational learning are building on sand. The through-line connecting both: the developers getting the most from AI are the ones who bring enough baseline knowledge to evaluate what the model gives them. The people struggling are the ones who treat output as finished work.

A third HN thread on the user frustration of coding agents noted that the real UX problem is that most users don't understand the context window is finite, that compaction is happening silently, and that the agent's 'memory' is not what they think it is. Taken together, these threads paint a picture of a community actively working out what a healthy relationship with AI-assisted development actually looks like.