Anthropic/YouTube/25/07/Claude Code Best Practices

• Claude Code best practices - YouTube - Claude/Code

  • My Notes

    • Nice reminder of Agentic definition from Anthropic/Blog/24/12/Building Effective Agents
      • Pure agent approach: instructions + powerful tools + model runs in loop
    • It's interesting to hear how skeptical he sounds of MCP use in Claude Code at this time; it sounds "advanced" and he advises that people prefer well-documented CLI/Tools over MCP Servers
      • • Install MCP/Servers for custom tools
          • Best Practice #Tip: Prefer well-documented CLI/Tool over MCP servers
    • I don't think I've managed to use the AI/Image/Processing features of Claude/Code yet
      • Screenshot Guidance 16:40

  • Video

    • • 00:05 Introduction and Personal Background

        • Welcome and Overview

          • "Welcome everyone to Claude Code best practices"
          • Talk structure:
            • High-level overview of what Claude Code is
            • Understanding how it works under the hood
            • Good use cases and best practices

        • Personal Introduction

          • Cal's background:
            • Joined Anthropic about a year and a half ago
            • Part of the Applied AI team
            • Mission: help customers build great products on top of Claude
            • Spends time prompting Claude for optimal outputs

          • Coding Background

            • Self-described coder who starts many projects but rarely finishes
            • Has a "graveyard of code" from unfinished projects
            • Always spinning up new ideas

      • 01:04 Discovery of Claude Code

        • Initial Encounter

          • Heard about new tool in Slack
          • Downloaded on a Friday night
          • Applied to a new notetaking app project

        • Transformative Experience

          • "That whole weekend just kind of totally changed the way that I code and think about software engineering"
          • Carried laptop around all weekend
          • "Super addicted to just watching Claude Code work"
          • Watched powerful application come together in real-time
          • Got much further than would have on own

        • Internal Recognition

          • Worried about token usage
          • Claude Code team had internal leaderboard tracking usage
          • Shot to the top over the weekend
          • Met Boris and Cat from early Claude Code team
          • Became core contributor focusing on:
            • System prompts
            • Tool descriptions and results
            • Evaluation methodology

      • 02:34 What is Claude Code?

        • Mental Model

          • "Claude Code is like that co-worker that does everything on the terminal"
          • Never touches GUI, terminal whiz
          • Like having a mentor (Tony) next to you all the time

        • Technical Architecture

          • Pure agent approach: instructions + powerful tools + model runs in loop
          • Tools include:
            • File creation and editing
            • Terminal usage
            • MCP (Model Context Protocol) integration
          • No indexing approach: Claude explores codebase like a new team member
          • Uses same search tools humans use: glob, grep, find
          • Agentic search: model can do searches, analyze results, then search more

        • UI and Security Features

          • Light UI layer showing Claude's work in real-time
          • Permission system for dangerous operations
          • Security focus with lightweight architecture
          • Easy integration with AWS and GCP cloud providers

      • 06:06 Use Cases for Claude Code

        • Discovery and Onboarding

          • Supercharges onboarding into new codebases
          • Ask "where is this feature implemented?"
          • Analyze git history and code evolution

        • Thought Partnership

          • Use as planning partner before diving into work
          • Ask for multiple implementation options
          • Validate approaches before starting

        • Code Building and Writing

          • Zero to one projects: build apps from empty directory
          • Existing codebases: primary focus area
          • High unit test coverage due to ease of adding tests
          • Great commit and PR messages

        • Deployment and Lifecycle

          • Use Claude Code SDK for headless automation
          • Integrate into CI/CD pipelines
          • GitHub Actions integration

        • Support and Scale

          • Faster debugging
          • Large codebase migrations (Java version upgrades, PHP to React)
          • CLI tool expertise (Git, Docker, BigQuery)

      • 10:17 Best Practices

        • Use Claude.md Files

          • 10:55 Claude Code doesn't really have AI/Memory
          • Main way to share state across sessions
          • Plopped into context when Claude Code starts
          • Can be placed in:
            • Project directory (shared with team) 11:28
            • Home directory (personal preferences)
          • Include information like:
            • How to run unit tests
            • Project layout overview
            • Style guides
            • Module locations

        • Permission Management 12:05

          • Read actions allowed automatically
          • Write/run commands require permission
          • Autoaccept mode: Shift+Tab for automatic approval Keyshort
          • Configure specific command approvals in settings - npm run test for example

        • Integration Setup 13:06

          • Leverage CLI tools (GitHub's gh tool)
          • Install MCP/Servers for custom tools
            • Best Practice #Tip: Prefer well-documented CLI/Tool over MCP servers
          • Document internal tools in Claude.md

        • Context Management

          • Context builds up over time (200,000 token limit) 14:21
          • Options when context fills up:
            • /clear: start over (preserves Claude.md)
            • /compact: summarize and continue 14:55
          • Compact functionality tuned for efficient continuation

      • 14:33 Efficient Workflows

        • Planning and To-Dos

          • Ask for plans before implementation 15:26

          • "Can you search around, figure out what's causing it, and tell me a plan?"
            

            • you can verify what claude's going to do
          • Watch to-do lists and intervene if needed
            • Press Escape to modify plans mid-execution

        • Smart Iterative Coding 16:10

          • Test-driven development approach 16:22
          • Small changes with test runs
          • Check TypeScript and linting
          • Commit regularly for fallback points

        • Screenshot Guidance 16:40

          • Use screenshots for visual guidance
          • Paste images or reference image files
          • "Look at this mock.png and build the website"

        • Advanced Techniques 17:15

          • Multiple Claude instances: run 2-4 simultaneously
          • Use tmux or different tabs

            • I can only do 2, but I know people who can do four 17:30 #Quote
              

          • Escape key mastery: stop and interject when needed
          • #Tip Double escape: jump back in conversation 18:03
          • MCP Server: for advanced tool integration
            • hear him be a bit hesitant about this
          • Headless automation: programmatic Claude usage

      • 18:35 New Features and Updates

        • Model Selection

          • /model command to see current model
          • Switch between Sonnet and Opus
          • /config for configuration changes 19:14

        • Enhanced Thinking 19:37 #Tip

          • "Think hard" or extended thinking capabilities - you should specify it

          • 19:45 Starting with Anthropic/Model/Claude/4 Models now think between tool calls AI/Model/Reasoning
            

            • Visible as lighter gray text
            • Encouraged for complex tasks and bug solving

        • IDE Integrations 20:28

          • VSCode VS Code and JetBrains integrations
            • Claude knows what file you have open
            • Enhanced development workflow

        • Staying Updated about new things in Claude/Code 21:01

          • #Tip
            • Public GitHub project: Claude Code under Anthropic
            • Check changelog weekly
              • Fast shipping pace

              • "Even I can't keep up with it"
                

      • 21:33 Q&A Session

        • Multiple Claude.md Files 21:40

          • Can have multiple files in different directories
          • Only reads Claude.md in current working directory by default
          • Encourages Claude to read relevant files in subdirectories 22:10
            • Can reference other files with @ symbol in Claude.md
          • Example: @package.json

        • Claude/Code/Claude.md Effectiveness - Why doesn't it work? 22:52

          • I have not had luck having Claude honor Claude.md, for example, inline comments #Quote -- Claude User
            

            • Model behavior issues (excessive commenting) mostly fixed in Claude 4
            • Better instruction following in newer models
            • Good time to review and update Claude.md files

        • Multi-Agent Execution 24:20

          • question: can we have agents that use each other's context?
          • Currently single agent approach
          • Future: shared markdown files for agent communication
          • Manual orchestration through file-based state sharing

        • Closing

          • Claude Code stickers available
          • Happy to answer questions and do live demos