Quartz 5

The Complete Claude Code Tutorial

Properties1
tagstutorial, ai, claude-code

3 min read

Summary

A beginner-to-intermediate Claude Code playbook from a CTO/ex-Amazon engineer arguing that output quality is bounded by input quality. Core lessons: think and plan before typing (use plan mode), treat CLAUDE.md as a short, specific, “why-not-just-what” leverage file, manage context aggressively (quality degrades around 30%, not 100%), write specific prompts that also state what NOT to do, switch between Opus for planning and Sonnet for execution, experiment with MCP/hooks/slash commands, and ultimately build automated systems (headless -p mode) rather than one-off interactions.

一份由 CTO／前 Amazon 工程師撰寫的 Claude Code 入門到進階指南，核心論點是「輸出品質受限於輸入品質」。重點包括：動手前先思考並規劃（用 plan mode）、把 CLAUDE.md 當作簡短而具體、講「為什麼」而非只講「做什麼」的槓桿檔案、積極管理 context（品質約在 30% 而非 100% 時開始下降）、寫出明確且說明「不要做什麼」的提示、用 Opus 規劃並切到 Sonnet 執行、實驗 MCP／hooks／slash commands，最終建立自動化系統（headless -p 模式）而非一次性互動。

Key Points

  • Think first: plan mode (Shift+Tab twice) consistently beats just-start-talking; specify architecture, not just the desired output.
  • CLAUDE.md should be short (Claude reliably follows ~150–200 instructions, ~50 already used by the system prompt), project-specific, explain why, and updated via the # key.
  • Context degrades at ~20–40% usage; use external memory files (SCRATCHPAD.md/plan.md), scope one conversation per task, and the copy-paste reset (/compact → /clear → paste essentials).
  • Prompting is communication: be specific, give examples over descriptions, and tell Claude what NOT to do (it tends to overengineer).
  • Model handoff: Opus for reasoning/planning, Sonnet for clear execution; CLAUDE.md keeps both under the same constraints.
  • When stuck, change approach (clear, simplify, show-don’t-tell, reframe) instead of looping more corrections.
  • Build systems: headless -p mode enables scripting, automated PR reviews, and a flywheel of logged improvements.

Insights

The article’s spine — “output is everything, but it only comes from input” — reframes prompting as the bottleneck rather than model capability, which is a useful corrective to model-shopping. The context-degradation claim (quality dropping at ~30%, so compaction can’t rescue an already-degraded session) is the practically important and counterintuitive piece, and the “Claude is stateless, plan accordingly” mental model directly justifies bot_vault’s reliance on CLAUDE.md and external scratch files. Pairs naturally with the Level 2 tutorial.

Connections

Raw Excerpt

Bad CLAUDE.md looks like documentation written for a new hire. Good CLAUDE.md looks like notes you’d leave yourself if you knew you’d have amnesia tomorrow.

Graph View