Learn Claude Code by doing, not reading

Original Article ↗ Hacker News Discussion ↗

Purpose of a Claude Code Tutorial

  • Some argue a tutorial for a natural-language-based tool is unnecessary: “just tell it what you want” and, if needed, have one AI build translation layers for another.
  • Others counter that you must still learn:
    • What the tool can do (context windows, compaction, agents, tools, plugins, MCP servers).
    • What “proper” vs “improper” usage looks like, similar to learning any complex tool.
  • A few liken it to “how to use university” guides: the medium is natural language, but meta-skills and capabilities still need teaching.

Learning by Doing vs. Learning Wrong

  • Some recommend skipping tutorials and just installing Claude Code and experimenting.
  • Others worry “learning by doing” can yield working results but a completely wrong mental model of how the system behaves.
  • LLM non-determinism and “cheerful failure” (confidently wrong output) complicate systematic learning.

Quiz, Pedagogy, and Site Quality

  • Several users report the “find your level” quiz labeling them “Beginner” even with advanced answers.
  • One person inspects the frontend logic and finds a scoring bug that can misclassify results.
  • Some say a broken entry quiz undermines trust; others highlight the real value in the 11 interactive modules, terminal simulators, and config builders.

Costs, Tokens, and Context Windows

  • Many complain about rapidly consumed quotas, especially with Opus 4.6 and 1M-token context.
  • Token-based billing is described as opaque and unintuitive; several call for request-based or clearer cost estimates.
  • There are mentions of:
    • Hidden caching behavior and quadratic cost growth with large contexts.
    • Environment flags and model switches to avoid 1M context, but UX is seen as confusing.
  • Concerns about “enshittification,” reduced quotas, and perceived monetization-driven defaults (e.g., defaulting to expensive 1M context).

Attitudes Toward AI Tools and Future

  • Some are enthusiastic daily users (agents, subagents, plugins) and see this as an essential new skill.
  • Others are deeply skeptical, calling LLMs “non-deterministic black boxes,” over-marketed, and only marginally better than good prompting.
  • There is anxiety about job pressure to use AI, but also pushback that one can still work without it.
  • A minority urges focusing on timeless programming fundamentals, predicting today’s AI tools may become just another auxiliary tool later.