obsidian/01 Projects/Claude code.md

how to how to_en

Based on the provided video guide, here is detailed process documentation designed for training new users on how to effectively build applications using Claude Code.

---

Standard Operating Procedure: Application Development with Claude Code

1. Objective

To build high-quality, bug-free applications using Claude Code by avoiding generic prompts ("slop") and utilizing a structured, iterative development cycle.

2. Prerequisites

• Access to the Claude Code terminal interface.
• A basic plan file or concept for the application (e.g., "A meditation app").

---

3. The Process

Phase 1: Detailed Specification (The Interview Method)

Do not begin by asking Claude to "build the app" immediately. Instead, force the AI to identify constraints and technical requirements first.

Step 1.1: Initiate the Interrogation Open your terminal and input the following prompt exactly. This commands Claude to use the AskUserQuestionTool to interview you about the project.

Command:

> Read this plan file and interview me in detail using AskUserQuestionTool about literally anything: technical implementation, UI & UX, concerns, tradeoffs, etc.

Step 1.2: Answer Technical Inquiries Claude will pause and ask a series of specific questions to generate a specification. You must answer these to define the scope. Expect questions regarding:

• Technical Details: (e.g., "iOS, Android, or Web?", "React Native or Flutter?")
• Audio/Data Handling: (e.g., "Stream or download?", "Offline mode required?")
• Edge Cases: (e.g., "What happens if a phone call interrupts a session?", "Does the music loop?")
• Trade-offs: (e.g., "Monetization vs. User Experience")

Outcome: A fully fleshed-out technical blueprint that accounts for edge cases before any code is written.

---

Phase 2: Modular Development (Feature Segmentation)

Do not attempt to generate the entire codebase in a single prompt.

Step 2.1: Break Down Features Divide your application specification into distinct, isolated features.

• Example: Instead of "Meditation App," break it down into:
  • Feature A: Audio Library
  • Feature B: Session Timer
  • Feature C: Progress Dashboard

Step 2.2: Build and Test Sequentially Instruct Claude to build one feature at a time.

1. Build Feature A.
2. Test Feature A to ensure it works perfectly.
3. Only after Feature A is verified, move on to Feature B.

Rule: Never move to the next feature until the current feature is fully functional and tested.

---

Phase 3: Context Window Management

AI models degrade in performance and "forget" instructions as the conversation gets too long. You must actively manage the context window.

Step 3.1: Monitor Context Usage Frequently check your token usage by running the context command.

Command:

/context

Step 3.2: The 50% Rule Watch the "Context Usage" percentage.

• The Danger Zone: Although the model may support 200k tokens, instruction adherence drops significantly after 40% - 50% usage (approx. 80k - 100k tokens).
• Symptoms: The AI will start ignoring previous instructions, introducing bugs, or hallucinating code.

Step 3.3: Session Reset When you hit the 40-50% mark:

1. Stop the current session.
2. Start a new session.
3. Resume building the next feature (e.g., Feature C) in the fresh session.

---

Summary Checklist

1.  Prompt: Did you use the AskUserQuestionTool prompt to force a technical interview?
2.  Scope: Did you answer all questions regarding edge cases and tech stack?
3.  Segmentation: Are you building one isolated feature at a time?
4.  Context: Is your context usage below 50%? If not, have you started a new session?