How to Better Use Claude Code

A practical guide synthesized from presentations by Patrick Ellis and Anand Tyagi (2025).

What Claude Code Actually Is

Claude Code is more than code generation - it’s the first truly competent AI agent we get to interface with directly.

Key characteristics:

  • Claude fine-tuned/RL’d specifically for Claude Code tools
  • Access to tools: bash, file operations, web search, todo lists, sub-agents, MCP servers
  • Model switching via /model command (Sonnet 4 vs Opus 4)

The Agent Loop 

Get Task → Add to Task List → Do Task → Reflect on Output → Output

Each “Do Task” step involves tool calls: reading files, searching, executing commands, or MCP calls.

Types of AI Agents (Context)

Type Examples
Chat-based ChatGPT, Gemini, Claude Desktop
CLI/IDE Agent Claude Code, Cursor, Windsurf, Kiro, Cline
Background Agent Codex, Jules, Claude Code + GitHub Actions, Devin
Agent Swarm Factory, custom multi-agent workflows

What Agents Need for Great Performance

1. Context

The most important factor. Methods to provide context:

  • Planning Mode - Use prompting to think through problems
  • /context directory - Following Amazon Kiro’s pattern: requirements.md, design.md, task-list.md
  • MCPs - Playwright, GitHub, Context7 for live data
  • /add-dir command - Add entire directories to context
  • Sub-agents - Use to summarize and keep everything in context

2. Evals (Quality Controls)

  • Examples of good/bad outputs
  • Linters and standards
  • Acceptance criteria
  • Automated tests

3. Tools

  • MCP servers (local and remote)
  • Web search capabilities
  • Bash/shell access

CLAUDE.md Mastery

The main context file - it becomes part of every prompt Claude receives.

“Your CLAUDE.md files become part of Claude’s prompts, so they should be refined like any frequently used prompt.” - Claude Code Documentation

Setup Strategy

  1. Add file structure with descriptions

    • List major entry points for each file
    • Describe what each component does

  2. Create hierarchical CLAUDE.md files

    • Main CLAUDE.md at root
    • Sub-folder CLAUDE.md files for major directories
    • Claude will fetch relevant sub-folder files based on context

  3. Use memories - The /memory command helps add context more easily

CLAUDE.md Content Ideas

MCP Configuration:

## Available MCPs

- GitHub: For repository operations
- Playwright: For browser automation
- Context7: For documentation lookup

Design Principles:

  • Use Deep Research to generate design-principles.md
  • Reference external style guides (e.g., “Conventional Commits”, “Airbnb JS Style Guide”)

Other Useful .md Files

File Purpose
CHANGELOG.md Track major updates
plan.md Large project planning
Development notes Keep track of decisions

When to Use Claude Code vs. Other Tools

Use Claude Code For:

  • Multi-step processes
  • Starting new projects
  • Complex tasks
  • Long-running operations
  • Exploring/ramping up on codebases
  • Refactoring large files
  • Generating files requiring info from many sources (README, Postman collections)
  • Generating tests with feedback loops

Use Cursor/Other IDE Tools For:

  • Specific problems in specific files/lines
  • One-step tasks
  • Managed, incremental development
  • Fine-grained control over edits

Common Struggles and Solutions

Problem Solution
Managing edits and changes Use git, or trust the agent
“Abuses grep” - keeps searching for known things Well-structured CLAUDE.md files
Context window limits Use sub-agents to summarize

Essential Commands

Command Use Case
/model Switch between Sonnet 4 and Opus 4
/add-dir Add directory context
/memory Add memories to CLAUDE.md
/plan Enter planning mode
/clear Clear context

Planning and Sub-agents

Planning Mode

  • Use before complex tasks
  • Think through execution steps
  • Creates structured approach

Sub-agents

  • Spawn for specific sub-tasks
  • Summarize large contexts
  • Parallel execution capabilities

Multi-Agent Workflows

Git Worktrees for Parallel Development 

git worktree add <PATH> <BRANCH-NAME>
git worktree list

This enables running multiple Claude Code instances on different branches simultaneously.

Claude Code + GitHub Actions

  • Run up to 30 agents at once
  • Async deployment of Claude Code
  • CI/CD integration for autonomous work

Beyond Engineering

Claude Code can handle non-coding tasks:

  • Second Brain management
  • Computer Admin - naming screenshots, organizing files, pipe operations
  • YouTube summary + execution of tutorials
  • Image manipulation, 3D modeling
  • Building slide decks
  • Workflow automations (e.g., n8n + headless Claude Code)

MCP Servers

Types

Type Examples
Local Playwright, Sequential Thinking, browser-tools-mcp, Blender
Remote GitHub, FireCrawl, Sentry
Platform-specific Gemini Apps, Anthropic Partners
Custom For logic encapsulation

Go-To MCPs for Coding

  1. GitHub (or via CLI)
  2. Playwright - browser automation
  3. Context7 - documentation lookup
  4. FireCrawl - web scraping
  5. Markitdown - document conversion
  6. Notion - note integration

MCP Registries

The Future: Orchestration Mindset

“We need to get comfortable delegating. Build the system that ‘compiles’ into the code. Your ‘spec’ is your new ‘source code’.” - Patrick Ellis

Key shifts:

  • Think of prompts and specs as source code
  • Design for agent orchestration
  • Delegate confidently to sub-agents

Final Tips

  1. Use Claude as an independent agent - a “super sidekick”
  2. Plan before executing - use plan mode or external plan files
  3. Maximize context - CLAUDE.md, commands, MCPs
  4. Use GitHub integration for multi-agent, async deployment
  5. Keep everything in model context - use sub-agents to summarize when needed

Resources

Official Anthropic

Community Resources

Key People to Follow (X/Twitter)

Synthesized from presentations by Patrick Ellis (patrickellis.io) and Anand Tyagi, 2025.