AI Coding Tools Landscape
Цей контент ще не доступний вашою мовою.
AI/ML Engineering Track | Complexity:
[MEDIUM]| Time: 4-5
Or: How I Learned to Stop Typing and Let the AI Do It
Section titled “Or: How I Learned to Stop Typing and Let the AI Do It”Reading Time: 4-5 hours Prerequisites: Module 0 complete, a willingness to let go of your keyboard
What You’ll Be Able to Do
Section titled “What You’ll Be Able to Do”By the end of this module, you will:
- Understand the AI development landscape in 2024-2025
- Know the major AI coding assistants and their strengths/weaknesses
- Develop a mental model for AI pair programming
- Recognize when to use AI vs traditional coding approaches
- Set up and configure your AI coding environment
- Master AI coding best practices from day one
- Complete your first AI-assisted coding project
- Understand the limitations and ethical considerations
Why this matters: AI coding assistants are the most significant productivity tool since IDEs were invented. Developers who master them are 3-10x more productive. Those who don’t are being left behind.
2025 Update: The landscape has evolved dramatically! Terminal/CLI coding agents (OpenAI Codex CLI, Gemini CLI) are now the most powerful category. FREE tiers have improved significantly (Gemini CLI is excellent!). This module now covers 13 major tools with verified 2025 pricing.
The Productivity Revolution: Why This Changes Everything
Section titled “The Productivity Revolution: Why This Changes Everything”The Problem Everyone Faces
Section titled “The Problem Everyone Faces”Without AI assistants (the old way):
Monday morning: Need to build user authenticationHour 1: Google "python authentication best practices"Hour 2: Read Stack Overflow, 17 tabs openHour 3: Copy-paste code, doesn't work, debug syntax errorsHour 4: Finally get basic version workingHour 5: Write tests, discover edge case bugsHour 6: Refactor, add error handlingHour 7-8: Documentation, code review feedback, more fixes
Result: 8 hours for basic auth, exhaustedWith AI assistants (the new way):
Monday morning: Need to build user authenticationHour 1: "Create Python auth system with JWT, bcrypt, rate limiting" → AI generates complete implementation Review code, ask questions, understand approachHour 2: Run tests AI generated, fix one edge case Add to your codebase, customize for your needs "Add OAuth2 social login support" → AI adds it
Result: 2 hours for production-ready auth, energizedThe difference: 8 hours of grinding vs 2 hours of directing. 75% time saved.
Did You Know? The Hidden Impact of AI Coding
Section titled “Did You Know? The Hidden Impact of AI Coding”Real data from GitHub (2024):
- 55% of professional developers use AI coding assistants daily
- 40% of new code at companies using Copilot is AI-generated
- 87% faster task completion for developers who use AI effectively
- #1 skill gap: Knowing HOW to direct AI, not IF to use it
But here’s the paradox: Developers now write MORE code in less time, but spend MORE time on design and architecture. The bottleneck shifted from typing to thinking.
Good news: This module teaches you to think like an AI-augmented developer!
Introduction: The AI Development Revolution
Section titled “Introduction: The AI Development Revolution”Imagine having a brilliant coding partner who:
- Never gets tired or frustrated
- Has read millions of codebases across every language
- Can write boilerplate instantly while you focus on architecture
- Helps you debug at 3 AM without complaint
- Explains complex concepts clearly in seconds
- Suggests better approaches you hadn’t considered
- Generates comprehensive tests while you code features
This isn’t science fiction. It’s your reality starting today.
The way we write code fundamentally changed in 2022-2023. AI coding assistants went from “interesting experiment” to “can’t live without them.” By 2025, coding without AI feels like writing essays with a typewriter when you have a computer available.
But here’s the critical insight: AI doesn’t replace developers. It amplifies them.
The Personality: AI is your tireless apprentice - brilliant at execution, needs your wisdom for direction.
Real-world analogy: Using AI coding assistants is like conducting an orchestra. You don’t play every instrument (write every line), but you direct the performance (architecture, verification, integration).
The best developers in 2025 aren’t those who can write the most code. They’re those who can direct AI to write the right code, then verify, refine, and integrate it effectively.
You’re about to learn how to be that developer.
The AI Development Landscape (2024-2025)
Section titled “The AI Development Landscape (2024-2025)”The Three Waves of AI Coding
Section titled “The Three Waves of AI Coding”Wave 1: Autocomplete (2021-2022)
- GitHub Copilot launches
- AI suggests next lines as you type
- Like smart autocomplete on steroids
- Impact: 30-50% faster coding for boilerplate
Wave 2: Chat + Code (2023-2024)
- ChatGPT for coding, Claude, Gemini
- Conversational AI that explains and codes
- Can handle complex multi-step tasks
- Impact: 2-5x faster for new features, debugging
Wave 3: Agentic AI (2024-2025) ← We are here
- AI that can read your entire codebase
- Takes multi-step autonomous actions
- Edits multiple files, runs commands, tests code
- Impact: 5-10x faster for refactoring, migrations, complex tasks
You’ll learn to use all three waves together.
Major AI Coding Assistants (Complete Guide)
Section titled “Major AI Coding Assistants (Complete Guide)”The landscape has exploded with options. Here’s your decision matrix:
Category 1: IDE-Integrated Autocomplete
Section titled “Category 1: IDE-Integrated Autocomplete”The Personality: These are your speed typists - incredibly fast at completing what you start typing.
1. GitHub Copilot (The Pioneer - 2021)
Section titled “1. GitHub Copilot (The Pioneer - 2021)”What it is: AI pair programmer that suggests code as you type
The Good:
- Autocomplete on steroids - predicts entire functions
- Learns from your codebase and coding style
- Excellent for boilerplate and repetitive code
- Built into VS Code, JetBrains IDEs, Vim, Neovim
- Fast, sub-second suggestions
- Mature product, well-tested
The Not-So-Good:
- ️ Line-by-line suggestions (can’t refactor across files)
- ️ Can’t execute commands or run tests
- ️ No reasoning explanation (suggests but doesn’t explain WHY)
- ️ Limited context window (~8K tokens)
Best for:
- Writing individual functions
- Completing patterns you’ve started
- Generating unit tests
- Boilerplate code (getters, setters, CRUD operations)
Pricing: $10/month individual, $19/month business (free for students, open source maintainers)
Real-world example: You write def calculate_fibonacci(n): and Copilot suggests the entire implementation in 0.5 seconds.
2. Tabnine
Section titled “2. Tabnine”What it is: Privacy-focused AI code completion
The Good:
- Can run locally (on-device models, no cloud needed)
- Privacy-first - doesn’t send code to cloud by default
- Supports 50+ IDEs and editors
- Team training on your codebase
- Compliant with enterprise security (SOC 2, GDPR)
The Not-So-Good:
- ️ Local models less powerful than cloud-based
- ️ Limited chat/explanation features
- ️ Smaller community than Copilot
Best for:
- Enterprises with strict privacy requirements
- Developers who want on-device AI
- Teams wanting to train on proprietary codebases
Pricing: Free tier, $12/month Pro, custom Enterprise pricing
Real-world example: Finance company trains Tabnine on their internal trading system code - AI learns company-specific patterns without code leaving their network.
3. Codeium
Section titled “3. Codeium”What it is: Free AI code completion and chat (Copilot alternative)
The Good:
- Completely FREE for individuals (unlimited usage!)
- Autocomplete + chat interface (best of both)
- Multi-file context awareness
- 70+ programming languages supported
- No usage limits on free tier
The Not-So-Good:
- ️ Newer product, less mature
- ️ Smaller training dataset than Copilot
- ️ Some features still in beta
Best for:
- Students and learners (FREE!)
- Developers wanting to try AI coding without payment
- Teams on tight budgets
Pricing: Free for individuals, $12/user/month for teams
Real-world example: Bootcamp students use Codeium to learn faster without spending $10/month - generates React components while they learn the patterns.
Category 2: AI-First IDEs
Section titled “Category 2: AI-First IDEs”The Personality: These are AI-native workshops - built from ground up for AI-driven development.
4. Cursor (The Game-Changer - 2023) ⭐ MOST POPULAR AI-FIRST IDE
Section titled “4. Cursor (The Game-Changer - 2023) ⭐ MOST POPULAR AI-FIRST IDE”What it is: VS Code fork with AI deeply integrated throughout
The Good:
- Sees your entire project (reads all files for context)
- Can edit multiple files simultaneously
- Runs terminal commands autonomously
- Explains reasoning step-by-step
- Long context window (200K+ tokens with Claude)
- Agentic (takes multi-step actions without asking)
- Composer mode for complex refactoring
- Import all VS Code extensions
- Supports Claude, gpt-5, or bring your own API key (BYOK)
The Not-So-Good:
- ️ Requires switching from your current IDE
- ️ Monthly subscription + usage credits model
- ️ Can make mistakes (always verify!)
- ️ Some VS Code features lag behind
Best for:
- Complex refactoring across multiple files
- Debugging mysterious issues
- Architecture changes
- Learning new codebases
- Developers who want best-in-class AI IDE experience
Pricing (CORRECTED for 2025):
- Hobby: FREE (entry-level limits on Tab autocomplete and Agent)
- Pro: $20/month
- Unlimited Tab (autocomplete)
- Unlimited Auto
- $20/month credit pool for frontier models (charged at API pricing)
- Can add overages if you exceed credit pool
- Ultra: $200/month
- 20× more usage than Pro
- For power users
- Teams: $40/user/month
- All Pro features + SSO and admin controls
- Annual plans: 20% discount
Recent Changes (2025):
- June 2025: Replaced request caps with usage credit pool system
- August 2025: Auto (agentic coding) now contributes to monthly usage
- Hybrid model: flat monthly fee + included usage credits + optional overages
Real-world example: “Refactor this Express app to use TypeScript and add input validation to all endpoints” - Cursor edits 15 files, adds types, validation, tests in 5 minutes.
5. Windsurf (by Codeium)
Section titled “5. Windsurf (by Codeium)”What it is: VS Code-based IDE with “Cascade” agentic AI
The Good:
- Free tier available (huge advantage over Cursor)
- Agentic workflows (multi-step autonomous coding)
- Codebase-aware suggestions
- Modern, clean UI
- “Cascade” feature for complex tasks
The Not-So-Good:
- ️ Newer product (launched late 2024)
- ️ Smaller user base, fewer tutorials
- ️ Some rough edges being smoothed out
Best for:
- Developers wanting Cursor-like experience for free
- Teams evaluating AI-first IDEs
- Students learning agentic AI coding
Pricing: Free tier (limited usage), $10/month Pro
Real-world example: Student uses Windsurf free tier to build entire Flask API with authentication, database, and tests - would cost $5+ in Cursor API fees.
️ Category 3: Terminal & CLI-Based Coding Agents ⭐ 2025 BREAKTHROUGH
Section titled “️ Category 3: Terminal & CLI-Based Coding Agents ⭐ 2025 BREAKTHROUGH”The Personality: These are command-line wizards - autonomous coding agents that live in your terminal. The newest and most powerful category!
6. OpenAI Codex CLI ⭐ NEW 2025 (Game-Changing!)
Section titled “6. OpenAI Codex CLI ⭐ NEW 2025 (Game-Changing!)”What it is: Official OpenAI terminal coding agent (launched October 2025)
The Good:
- INCLUDED with ChatGPT Plus/Pro ($20/month) - NO additional API costs!
- Zero-setup:
npm install -g @openai/codex - IDE extensions available (VSCode, Cursor, Windsurf)
- Uses GPT-5.1-Codex-Max (most advanced OpenAI coding model)
- Works across millions of tokens (project-scale refactors)
- 95% of OpenAI engineers use it weekly
- ReAct loop for complex multi-step tasks
- Available in CLI, IDE extensions, and web
The Not-So-Good:
- ️ Requires ChatGPT Plus/Pro subscription
- ️ Terminal-focused (but IDE extensions available)
- ️ Newer tool (launched Oct 2025, still maturing)
Best for:
- Developers already paying for ChatGPT Plus ($20/month gets you BOTH chat + CLI agent!)
- Large-scale refactoring (millions of tokens context)
- Multi-step autonomous coding tasks
- Terminal-first workflows
Pricing: INCLUDED with ChatGPT Plus ($20/month), Pro ($200/month), Business, Edu, Enterprise - No additional API costs!
Real-world impact:
- Used by Duolingo, Vanta, Cisco, Rakuten
- Engineers ship 70% more PRs after adopting Codex
- “Migrate entire Python 2 codebase to Python 3 across 200 files” - Codex CLI works autonomously for hours
This is huge! OpenAI Codex CLI is the ONLY major coding tool included in a chat subscription. ChatGPT Plus now gives you both unlimited web chat AND a powerful terminal coding agent!
7. Google Gemini CLI ⭐ FREE TIER (Best Free Option 2025!)
Section titled “7. Google Gemini CLI ⭐ FREE TIER (Best Free Option 2025!)”What it is: Open-source terminal AI agent from Google (launched June 2025)
The Good:
- FREE tier with generous limits (just login with Google account!)
- Gemini 2.5 Pro with 1 million token context window (largest available!)
- 60 requests/minute, 1,000 requests/day on free tier
- Open source (70,000+ GitHub stars, 2,800+ community PRs)
- ReAct loop with tool use
- MCP server support
- Integrated with Gemini Code Assist (Agent Mode in VS Code)
The Not-So-Good:
- ️ CLI UX still improving (community feedback: rougher than competitors)
- ️ Newer than established tools (launched June 2025)
- ️ Some features still in development
Best for:
- Budget-conscious developers (best free tier available in 2025!)
- Long context tasks (1M tokens beats most competitors)
- Google ecosystem users
- Open source enthusiasts
- Students and learners
Pricing:
- Free: Gemini Code Assist license (2.5 Pro, 1M context, 60 req/min, 1K req/day)
- Standard: Paid tier for higher limits
- Enterprise: Enterprise features and support
Real-world example: “Analyze this 500-file codebase and find all security vulnerabilities” - Gemini CLI loads entire project in 1M context window, systematically scans.
Note: User mentioned “Gemini CLI sucks” - but the free tier is actually excellent for most use cases! CLI UX is improving rapidly with community contributions.
8. Aider.ai (The Terminal Power User’s Dream)
Section titled “8. Aider.ai (The Terminal Power User’s Dream)”What it is: AI pair programming in your terminal with git integration
The Good:
- Works with ANY editor (terminal-based, editor-agnostic)
- Git integration (auto-commits AI changes)
- Can edit multiple files
- Uses gpt-5, Claude, or local models
- Perfect for SSH/remote workflows
- Scriptable for automation
- Open source
The Not-So-Good:
- ️ Terminal-only (no GUI)
- ️ Requires comfort with command line
- ️ No inline autocomplete
Best for:
- Terminal power users (Vim, Emacs users)
- Automated workflows and CI/CD
- Remote development over SSH
- Git-integrated AI coding
Pricing: Free (bring your own API key)
Website: https://aider.chat/
Real-world example: DevOps engineer SSHed into production server, uses Aider to debug and fix issue with AI assistance, all changes auto-committed with descriptive messages.
9. Cline (formerly Claude Dev)
Section titled “9. Cline (formerly Claude Dev)”What it is: VS Code extension for agentic AI coding with terminal access
The Good:
- Autonomous coding agent (can work for minutes unsupervised)
- Can create files, run commands, read output
- Works with Claude, gpt-5, local models
- Browser automation capabilities
- Integrated with VS Code but CLI-like power
The Not-So-Good:
- ️ Can run expensive operations without asking (watch your API costs!)
- ️ Sometimes over-engineers solutions
- ️ Requires trusting AI with filesystem access
Best for:
- Agentic workflows within VS Code
- Setting up new projects
- Repetitive migrations
Pricing: Free extension (bring your own API key)
Real-world example: “Set up a new Next.js project with Tailwind, ESLint, Jest, and create a basic auth flow” - Cline creates 20+ files, installs packages, configures everything in 3 minutes.
9b. Goose by Block (Square) (Full Automation Champion)
Section titled “9b. Goose by Block (Square) (Full Automation Champion)”What it is: Open-source AI agent for full software development automation from Block (formerly Square)
The Good:
- Fully autonomous - can work for extended periods without human intervention
- Desktop app + CLI interface (best of both worlds)
- Native MCP support (Model Context Protocol)
- Works with Ollama for local LLMs (Llama, Mistral, etc.)
- Session persistence (pick up where you left off)
- Can manage entire development workflows
- Backed by Block/Square (enterprise credibility)
- Open source (MIT license, GitHub: block/goose)
The Not-So-Good:
- ️ Newer project, rapidly evolving API
- ️ Less documentation than mature tools
- ️ Can be aggressive with file changes (always use version control!)
- ️ Requires careful permission management
Best for:
- Full project automation (“build me a REST API”)
- Developers who want hands-off AI agents
- MCP-integrated workflows
- Teams wanting open-source full automation
Pricing: Free (open-source, bring your own API key or use Ollama)
Website: https://github.com/block/goose
Real-world example: “Set up a complete microservice with FastAPI, PostgreSQL, Redis caching, Docker compose, and CI/CD pipeline” - Goose builds entire infrastructure in 15 minutes.
9c. Open Interpreter (Code Execution Specialist)
Section titled “9c. Open Interpreter (Code Execution Specialist)”What it is: Natural language interface to your computer’s code interpreter
The Good:
- Runs code locally in Python, JavaScript, shell, and more
- Direct computer control (files, applications, terminal)
- Works with Ollama for local LLMs
- Streaming output (see results as they happen)
- Voice interface available (talk to your computer!)
- Open source, active community
The Not-So-Good:
- Limited MCP support (not designed for it)
- ️ Security concerns (gives AI direct computer access)
- ️ Can execute destructive commands if not careful
- ️ Less focused on coding, more on general computer tasks
Best for:
- Data analysis and visualization
- System administration tasks
- Quick scripting and automation
- Prototyping ideas rapidly
Pricing: Free (open-source, bring your own API key or use Ollama)
Website: https://github.com/OpenInterpreter/open-interpreter
Real-world example: “Download the last 30 days of logs from S3, parse them for errors, create a visualization, and email the report” - Open Interpreter handles the entire pipeline.
MCP-Enabled Tools: The 2025 Game Changer
Section titled “MCP-Enabled Tools: The 2025 Game Changer”What is MCP (Model Context Protocol)?
Section titled “What is MCP (Model Context Protocol)?”MCP is Anthropic’s open standard for connecting AI assistants to external tools and data sources. Think of it as “USB for AI” - a universal way for AI tools to plug into your development environment.
Why MCP matters:
- File system access: AI can read/write project files
- ️ Database integration: Direct access to PostgreSQL, SQLite, etc.
- API connections: AI can call external services
- ️ Custom tools: Build your own MCP servers for domain-specific tasks
- Consistency: Same interface across different AI tools
MCP-Enabled Tools Comparison Matrix
Section titled “MCP-Enabled Tools Comparison Matrix”| Tool | Interface | MCP Support | Local LLM | Best For |
|---|---|---|---|---|
| Cline | VS Code + CLI | Native | Ollama | Agentic workflows |
| Continue.dev | IDE + CLI | Native | Ollama | Customizable, enterprise |
| Goose | Desktop + CLI | Native | Ollama | Full automation |
| Aider | CLI only | ️ Via server | Ollama | Terminal-first, Git focus |
| Open Interpreter | CLI only | Limited | Basic | Code execution |
Deep Dive: Each Tool’s MCP Integration
Section titled “Deep Dive: Each Tool’s MCP Integration”Cline - MCP Native Champion
Section titled “Cline - MCP Native Champion”MCP Capabilities:├── File Operations: Full read/write access├── Terminal Commands: Execute shell commands├── Browser Control: Web automation via Puppeteer├── Custom Servers: Add any MCP server└── Configuration: .cline/mcp_servers.jsonMCP Setup in Cline:
{ "mcpServers": { "filesystem": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/project"] }, "postgres": { "command": "npx", "args": ["-y", "@modelcontextprotocol/server-postgres"], "env": { "POSTGRES_URL": "postgresql://localhost/mydb" } } }}Continue.dev - Enterprise MCP Powerhouse
Section titled “Continue.dev - Enterprise MCP Powerhouse”MCP Capabilities:├── Customizable Tools: Define custom functions├── Context Providers: Inject data from anywhere├── Multiple Models: Route different tasks to different LLMs├── Local Models: Full Ollama integration└── Configuration: config.jsonMCP Setup in Continue.dev:
{ "experimental": { "mcpServers": [ { "name": "my-database", "command": "node", "args": ["./mcp-servers/postgres-server.js"] } ] }}Goose - Autonomous MCP Agent
Section titled “Goose - Autonomous MCP Agent”MCP Capabilities:├── Session Management: Long-running tasks├── Multi-step Plans: AI-driven orchestration├── Desktop App: User-friendly GUI├── Tool Chaining: Combine multiple MCP servers└── Configuration: ~/.goose/config.yamlMCP Setup in Goose:
mcp_servers: - name: github command: npx args: ["-y", "@modelcontextprotocol/server-github"] env: GITHUB_TOKEN: ${GITHUB_TOKEN}Aider - Limited MCP Support ️
Section titled “Aider - Limited MCP Support ️”Aider focuses on Git integration rather than MCP. However, you can add MCP-like capabilities via external servers:
# Aider doesn't have native MCP, but you can:# 1. Use Aider for Git-focused workflows# 2. Combine with MCP servers via shell commandsaider --model anthropic/claude-4.6-sonnetOpen Interpreter - Minimal MCP
Section titled “Open Interpreter - Minimal MCP”Open Interpreter takes a different approach - direct code execution rather than MCP protocols:
from interpreter import interpreter
interpreter.chat("Create a bar chart of monthly sales")# → Actually runs Python, creates the chart!Did You Know? The MCP Origin Story
Section titled “Did You Know? The MCP Origin Story”November 2024: Anthropic announced MCP (Model Context Protocol) as an open standard. Within 3 months:
- 500+ MCP servers were created by the community
- All major AI coding tools added MCP support
- Enterprise adoption exploded (companies can now connect AI to internal systems safely)
The insight that led to MCP: Anthropic engineers noticed that 80% of AI coding tool limitations came from context access, not model capability. An AI that can’t read your database schema will always give generic advice. MCP solves this.
Why this matters for Kaizen: Your Lean DevOps platform can expose internal APIs via MCP servers. AI agents can then:
- Query your metrics database
- Check deployment status
- Access internal documentation
- Trigger workflows (with approval gates)
The future: MCP is becoming the “language” that AI tools use to talk to your infrastructure. Learning it now puts you ahead of the curve!
Choosing Your MCP Tool Stack
Section titled “Choosing Your MCP Tool Stack”For Kaizen Development (recommended):
Primary: Cline or Continue.dev├── MCP servers for: Postgres, GitHub, filesystem├── Local LLM: Ollama with codestral/deepseek-coder├── Cloud LLM: Claude 3.5 Sonnet for complex reasoning└── Workflow: Git-based with PR reviewsFor Rapid Prototyping:
Primary: Goose├── Full automation mode├── Session persistence for multi-day projects└── Desktop app for visibilityFor Data Analysis:
Primary: Open Interpreter├── Direct Python execution├── Visualization output└── Great for exploring data before productionizingCategory 4: Agentic Assistants (Extensions)
Section titled “Category 4: Agentic Assistants (Extensions)”The Personality: These are autonomous agents - they can reason, plan, and execute multi-step tasks.
10. Anthropic Claude Code CLI ⭐ WHAT YOU’RE USING RIGHT NOW
Section titled “10. Anthropic Claude Code CLI ⭐ WHAT YOU’RE USING RIGHT NOW”What it is: Official Anthropic terminal coding agent - the AI assistant you’re using right now to learn!
The Good:
- Agentic multi-file editing
- Git workflow integration
- Understands entire codebase (200K token context)
- Uses Claude Opus 4.1, Sonnet 4.5, Haiku 3.5
- Works on macOS, Linux, Windows
- Open source (GitHub: anthropics/claude-code)
- No backend server needed (runs locally)
- Excellent at explaining and teaching (best for learning!)
The Not-So-Good:
- ️ Requires Claude API key (pay-per-use)
- ️ NOT included with Claude Pro subscription ($20/month web chat subscription does NOT give you API access!)
- ️ Standard API pricing (can add up with heavy use)
- ️ Terminal-only (no GUI)
Best for:
- Complex multi-file refactoring
- Git-integrated workflows
- Learning (Claude explains WHY, not just WHAT)
- Developers who prefer Claude models
- Terminal power users
Pricing:
- API-based: Standard Claude API pricing
- Sonnet 4.5: $3/M input tokens, $15/M output tokens
- Opus 4.1: More expensive (check current pricing)
- Haiku 3.5: Cheapest option
- ~$3-5 for this entire Neural Dojo curriculum
- ️ IMPORTANT: Claude Pro ($20/month) subscription only gives you web chat access, NOT API access for Claude Code CLI!
Real-world example: “Read this GitHub issue, write code to fix it, run tests, commit with descriptive message, open PR” - Claude Code handles entire workflow.
Why you’re using this for Neural Dojo: Claude excels at explaining concepts while coding - perfect for learning!
Real-world example: “Debug why my React app is re-rendering too much” - Claude reads component tree, finds memo() bugs, explains React rendering lifecycle, fixes issue across 4 files.
11. Continue.dev (The Open Source Champion)
Section titled “11. Continue.dev (The Open Source Champion)”What it is: Open-source AI code assistant for VS Code/JetBrains
The Good:
- Fully open source (MIT license)
- Works with ANY LLM (Claude, gpt-5, Llama, Mistral, local models)
- Autocomplete + chat modes
- Fully customizable (edit prompts, add tools)
- Privacy-friendly (can use local models)
- Active community, frequent updates
The Not-So-Good:
- ️ Requires more manual configuration
- ️ UI less polished than commercial tools
- ️ Some features still experimental
Best for:
- Developers wanting open-source solution
- Custom LLM setups (local models, custom APIs)
- Teams with specific privacy requirements
- Hackers who want to customize everything
Pricing: Free (bring your own API key or local model)
Website: https://continue.dev/
Real-world example: Startup runs Continue.dev with self-hosted Llama 4 70B model - zero API costs, all code stays internal, customized prompts for their domain-specific patterns.
Category 5: General AI for Coding
Section titled “Category 5: General AI for Coding”The Personality: These are versatile mentors - not specialized for code, but excellent at it.
12. ChatGPT (OpenAI)
Section titled “12. ChatGPT (OpenAI)”What it is: General-purpose AI that’s excellent for coding tasks
The Good:
- ChatGPT Plus ($20/month) or free tier
- Canvas mode for iterative code editing
- Code interpreter (runs Python in sandbox)
- Great for learning, debugging, code review
- No IDE integration needed (browser-based)
- Huge knowledge base, excellent explanations
- O1 model for complex reasoning
The Not-So-Good:
- ️ No direct IDE integration (copy-paste workflow)
- ️ Can’t see your full codebase
- ️ Can’t edit files directly
- ️ Context resets between conversations
Best for:
- Ad-hoc coding help
- Learning new concepts
- Debugging specific functions
- Code review outside your editor
- Algorithm design
Pricing: Free tier or $20/month Plus ($200/month Pro for O1 access)
Real-world example: Stuck on a complex algorithm, paste it into ChatGPT: “Explain this binary tree traversal and suggest optimizations” - get detailed explanation with visualizations, complexity analysis, and 3 alternative approaches.
13. Gemini Advanced (Google Web)
Section titled “13. Gemini Advanced (Google Web)”What it is: Google’s AI assistant with strong coding capabilities
The Good:
- Free tier available (generous limits)
- Integrated with Google Workspace
- Long context window (2M tokens in Pro!)
- Code execution capabilities
- Multimodal (can analyze images, diagrams, screenshots)
- Google Search integration
The Not-So-Good:
- ️ No IDE integration
- ️ Less popular for coding than ChatGPT/Claude
- ️ Copy-paste workflow
Best for:
- Those in Google ecosystem
- Long-context code analysis (entire files/repos)
- Multimodal tasks (analyzing architecture diagrams)
- Free tier users
Pricing: Free tier, $20/month for Advanced
Website: https://gemini.google.com/
Real-world example: Upload architecture diagram screenshot, ask “Convert this system design to a Python microservices implementation” - Gemini analyzes image, generates 5 services with gRPC communication, Docker compose, and deployment docs.
CRITICAL: Subscriptions vs API Access (Read This Carefully!)
Section titled “CRITICAL: Subscriptions vs API Access (Read This Carefully!)”The #1 Confusion in AI Coding (90% of Beginners Get This Wrong!)
Section titled “The #1 Confusion in AI Coding (90% of Beginners Get This Wrong!)”Question: “I have ChatGPT Plus. Can I use it with Aider.ai?” Answer: NO (Except: OpenAI Codex CLI IS included!)
Question: “I have Claude Pro. Can I use it with Claude Code CLI?” Answer: NO
Question: “I have Gemini Advanced. Can I use it with Continue.dev?” Answer: NO
Question from user: “Gemini CLI sucks, can I use my Gemini Advanced subscription instead?” Answer: NO (They’re separate products!)
Why This Module Matters Understanding the Two Separate Ecosystems
Section titled “Why This Module Matters Understanding the Two Separate Ecosystems”Ecosystem 1: Chat Subscriptions (Web Interface Only)
Section titled “Ecosystem 1: Chat Subscriptions (Web Interface Only)”| Product | Price | What You Get | Can Use With Coding Tools? |
|---|---|---|---|
| ChatGPT Plus | $20/mo | Unlimited web chat + Codex CLI | No (except Codex CLI) |
| Claude Pro | $20/mo | 5× more web chat usage | No - web only |
| Gemini Advanced | $20/mo | Web chat + Google Workspace | No - web only |
What subscriptions give you:
- Unlimited access to web chat interface
- Click buttons in a browser to get AI help
- Great for ad-hoc questions, learning, brainstorming
- Cannot be used programmatically by external tools
- Separate billing from API access
Exception: ChatGPT Plus/Pro DOES include OpenAI Codex CLI! (Only subscription that includes a coding tool)
Ecosystem 2: API Access (For Coding Tools)
Section titled “Ecosystem 2: API Access (For Coding Tools)”| Product | Pricing Model | What You Get | Can Use With Coding Tools? |
|---|---|---|---|
| OpenAI API | Pay-per-use | Programmatic access to GPT models | Yes - Aider, Continue.dev, Cursor BYOK |
| Anthropic API | Pay-per-use | Programmatic access to Claude models | Yes - Claude Code CLI, Aider, etc. |
| Google AI API | Free tier + paid | Programmatic access to Gemini | Yes - Continue.dev, custom tools |
What API access gives you:
- Your code/tools can call AI models
- Works with terminal agents (Aider.ai, Claude Code CLI)
- Works with IDE extensions (Continue.dev, Cursor BYOK)
- Pay per request (typically $3-20/month for moderate use)
- ️ Completely separate from chat subscriptions
The Confusion Explained (With Examples)
Section titled “The Confusion Explained (With Examples)”Scenario 1: You have ChatGPT Plus ($20/month)
What you CAN do:
- Use ChatGPT web interface for coding help
- Use OpenAI Codex CLI (INCLUDED!)
- Copy-paste code for debugging/review
What you CANNOT do:
- Use Aider.ai (needs separate OpenAI API account)
- Use Continue.dev with gpt-5 (needs separate API key)
- Use Cursor BYOK with gpt-5 (needs API key)
To use those tools: Sign up for OpenAI API separately at platform.openai.com (different from chat.openai.com!)
Scenario 2: You have Claude Pro ($20/month)
What you CAN do:
- Use Claude web interface (claude.ai) for coding help
- 200K token context for code analysis
- Artifacts for code editing
What you CANNOT do:
- Use Claude Code CLI (needs separate Anthropic API key)
- Use Aider.ai with Claude (needs API key)
- Use Continue.dev with Claude (needs API key)
To use Claude Code CLI: Get separate API key from console.anthropic.com (pay-per-use, ~$3-10/month)
Scenario 3: You have Gemini Advanced ($20/month)
What you CAN do:
- Use Gemini web interface for coding help
- 2M token context window
- Google Workspace integration
What you CANNOT do:
- Use Gemini CLI (it’s FREE but separate - just login with Google!)
- Use Continue.dev with Gemini (needs Google AI API key)
- Use custom integrations
Good news: Gemini CLI has an excellent FREE tier! Just login with your Google account at ai.google.dev
Can You Have Both? (Subscription + API)
Section titled “Can You Have Both? (Subscription + API)”YES! Many developers have:
- ChatGPT Plus for daily ad-hoc help ($20/month)
- Claude API for Claude Code CLI (~$5-10/month usage)
- Total: ~$25-30/month for complete AI coding setup
Or even better in 2025:
- ChatGPT Plus ($20/month) → includes Codex CLI!
- Gemini CLI (FREE tier) → for long context tasks
- Total: $20/month, two powerful terminal agents!
Cost Comparison by Use Case
Section titled “Cost Comparison by Use Case”“I’m a student on tight budget”:
- Skip subscriptions
- Use Gemini CLI (FREE tier - 1K req/day!)
- Use Codeium (FREE autocomplete)
- Use Continue.dev + Google AI API (FREE tier)
- Cost: $0/month
“I want best experience, budget not an issue”:
- ChatGPT Plus ($20/mo) → includes Codex CLI
- Cursor Pro ($20/mo) → best AI IDE
- Claude API (~$10/mo) → for Claude Code CLI when needed
- Cost: $50/month, every tool available
“I’m a professional developer, moderate budget”:
- Cursor Pro ($20/mo) → all-in-one: autocomplete + agentic AI + IDE
- Gemini CLI (FREE) → for long context tasks
- Cost: $20/month
“I prefer terminal/CLI workflows”:
- ChatGPT Plus ($20/mo) → includes Codex CLI
- Gemini CLI (FREE)
- Use your favorite editor (Vim, Emacs, VS Code)
- Cost: $20/month
Common Mistakes to Avoid
Section titled “Common Mistakes to Avoid”Mistake #1: “I paid for Claude Pro, so I can use Claude Code CLI”
- Reality: Claude Pro = web chat only, separate from API
- Fix: Get Claude API key from console.anthropic.com (~$3-10/month usage)
Mistake #2: “Gemini CLI sucks, I’ll use my Gemini Advanced subscription”
- Reality: Gemini Advanced = web only, can’t help with CLI
- Fix: Gemini CLI is FREE with Google account - try the free tier! UX improving rapidly
Mistake #3: “I’ll buy ChatGPT Plus and use Aider.ai with it”
- Reality: ChatGPT Plus doesn’t give API access for Aider.ai
- Fix: Use included Codex CLI instead, OR get separate OpenAI API account
Mistake #4: “All $20/month subscriptions are the same”
- Reality: They’re for different things:
- ChatGPT Plus = web chat + Codex CLI (unique!)
- Claude Pro = web chat only
- Gemini Advanced = web chat only
- Cursor Pro = AI IDE (not a chat service)
Bottom Line
Section titled “Bottom Line”Three separate product ecosystems - don’t confuse them:
-
Chat Subscriptions ($20/month each)
- ChatGPT Plus, Claude Pro, Gemini Advanced
- For web-based AI assistance
- Exception: ChatGPT Plus includes Codex CLI!
-
API Access (pay-per-use or free tiers)
- OpenAI API, Anthropic API, Google AI API
- For terminal tools (Claude Code CLI, Aider.ai, Continue.dev)
- Typically $0-20/month depending on usage
- Gemini has generous FREE tier!
-
IDE Subscriptions ($10-200/month)
- Cursor, GitHub Copilot, Windsurf
- For integrated IDE experiences
- Some allow BYOK to reduce costs
Want coding tools? → You need #2 (API Access) or #3 (IDE subscriptions) Want unlimited chat? → You need #1 (Chat Subscriptions) Want both? → Get both (separate billing, ~$20-50/month total)
Can’t afford both? → Start with API access or free tiers ($0-10/month), skip chat subscriptions
️ The Landscape Map (Visual Guide - 2025)
Section titled “️ The Landscape Map (Visual Guide - 2025)”AUTOCOMPLETE FOCUS ←──────────────────────→ AGENTIC FOCUSIDE-Integrated ←──────────────────────→ Terminal/Standalone
Copilot/Tabnine Cursor/Windsurf Codex CLI/Gemini CLI ChatGPT/Gemini AdvancedCodeium Gemini Code Assist Aider.ai/Cline Claude Web │ │ │ │ ├─ Line-by-line ├─ Hybrid ├─ Terminal agents ├─ General AI ├─ Fast (<1s) ├─ Multi-file ├─ Git-aware ├─ Long context ├─ Autocomplete ├─ Chat+code ├─ Multi-step tasks ├─ Reasoning ├─ Limited context ├─ Agentic ├─ Autonomous ├─ Teaching └─ $0-20/mo └─ $20-200/mo └─ $0-20/mo └─ Free-$20/mo
NEW 2025 Breakthrough: Terminal/CLI Coding Agents- OpenAI Codex CLI (included with ChatGPT Plus!)- Gemini CLI (FREE tier with 1M context!)- Claude Code CLI (API-based)
BYOK = Bring Your Own (API) KeySpeed vs Power Trade-off:
Copilot ─────────────────→ (speed) Fast suggestions, limited power
Claude Code/Cursor ──────→ (power) Slower, but can refactor entire codebasesCategories at a Glance
Section titled “Categories at a Glance”1. Autocomplete-First (Copilot, Tabnine, Codeium)
Section titled “1. Autocomplete-First (Copilot, Tabnine, Codeium)”- Best for: Fast, inline suggestions as you type
- Speed: (sub-second)
- Power: ⭐⭐ (single function)
- Learning curve: Minimal
- Cost: $0-20/month
2. AI-First IDEs (Cursor, Windsurf)
Section titled “2. AI-First IDEs (Cursor, Windsurf)”- Best for: Complete AI-integrated development environment
- Speed: (seconds to minutes)
- Power: ⭐⭐⭐⭐⭐ (entire project)
- Learning curve: Moderate (new IDE)
- Cost: $10-20/month
3. Terminal/CLI (Aider.ai, Cline)
Section titled “3. Terminal/CLI (Aider.ai, Cline)”- Best for: Automated workflows, git integration, power users
- Speed: (seconds to minutes)
- Power: ⭐⭐⭐⭐ (multiple files, git)
- Learning curve: Moderate-High (CLI comfort needed)
- Cost: BYOK (~$3-10/month in API)
4. Agentic Extensions (Claude Code, Continue.dev)
Section titled “4. Agentic Extensions (Claude Code, Continue.dev)”- Best for: Complex multi-file tasks, learning, architecture
- Speed: (seconds to minutes)
- Power: ⭐⭐⭐⭐⭐ (entire codebase)
- Learning curve: Low-Moderate
- Cost: BYOK (~$3-10/month in API)
5. General AI (ChatGPT, Gemini)
Section titled “5. General AI (ChatGPT, Gemini)”- Best for: Ad-hoc help, learning, no IDE setup needed
- Speed: (instant responses)
- Power: ⭐⭐⭐ (single task)
- Learning curve: Minimal
- Cost: Free or $20/month
The Modern Developer Stack (2025 Recommendations)
Section titled “The Modern Developer Stack (2025 Recommendations)”Best Overall Combo (If Budget Allows)
Section titled “Best Overall Combo (If Budget Allows)”Option A: All-in-One
- Cursor Pro: $20/month (IDE + autocomplete + agentic AI)
- ChatGPT Plus: $20/month (web chat + Codex CLI included!)
- Total: $40/month
- What you get: Complete AI coding setup, every major tool covered
Option B: Premium Everything
- Cursor Ultra: $200/month (20× more usage)
- ChatGPT Pro: $200/month (unlimited o1, o3-mini)
- Total: $400/month
- What you get: Unlimited everything, maximum productivity
Best Value Combo (Budget-Conscious)
Section titled “Best Value Combo (Budget-Conscious)”FREE Tier (Student/Learner):
- Gemini CLI: FREE (1M context, 1K req/day)
- Codeium: FREE (unlimited autocomplete)
- ChatGPT Free: FREE (ad-hoc help)
- VS Code: FREE
- Total: $0/month
- What you get: Surprisingly powerful setup with zero cost!
Budget Pro ($20/month):
- ChatGPT Plus: $20/month (includes Codex CLI!)
- Gemini CLI: FREE (long context tasks)
- Codeium: FREE (autocomplete)
- Total: $20/month
- What you get: Two terminal agents (Codex + Gemini) + autocomplete
Best for Learning (Neural Dojo Recommendation)
Section titled “Best for Learning (Neural Dojo Recommendation)”Primary: VS Code + Claude Code CLI (pay-per-use)
- Explains reasoning (critical for learning!)
- Can read entire project
- This curriculum uses Claude Code in examples
- Cost: ~$3-5 for entire Neural Dojo curriculum
- Why: Claude excels at teaching WHY, not just WHAT
Supplementary FREE tools:
- Gemini CLI: FREE (1M context) - for long codebase analysis
- Codeium: FREE - fast autocomplete while learning
- ChatGPT Free: FREE - quick questions
Total: ~$3-10/month (mostly Claude API usage)
Alternative if you already have ChatGPT Plus:
- ChatGPT Plus: $20/month (includes Codex CLI!)
- Gemini CLI: FREE
- Total: $20/month, covers most needs
Choosing Your Stack: Decision Tree (2025 Edition)
Section titled “Choosing Your Stack: Decision Tree (2025 Edition)”START: What's your budget? │ ├─ $0/month (FREE) ────────────────────────────┐ │ └─ Gemini CLI (FREE, 1M context) │ │ + Codeium (FREE autocomplete) │ ← Best free setup 2025! │ + Continue.dev (FREE) │ │ = Powerful AI coding, zero cost │ │ │ ├─ ~$20/month ──────────────────────────────────┤ │ ├─ Want terminal agents? │ │ │ └─ ChatGPT Plus ($20) → includes Codex CLI! │ │ + Gemini CLI (FREE) │ │ │ = Two CLI agents for $20/month │ │ │ │ │ └─ Want AI-first IDE? │ │ └─ Cursor Pro ($20) │ │ + Gemini CLI (FREE) │ │ = Best all-in-one experience │ │ │ ├─ ~$40/month (Recommended) ────────────────────┤ │ └─ Cursor Pro ($20) + ChatGPT Plus ($20) │ │ + Gemini CLI (FREE) │ │ = Complete setup: IDE + agents + chat │ │ │ └─ $200+/month (Ultimate) ──────────────────────┘ └─ Cursor Ultra ($200) or ChatGPT Pro ($200) = Unlimited everything
LEARNING (Neural Dojo students)? └─ VS Code (FREE) + Claude Code CLI (~$3-5 for curriculum) + Gemini CLI (FREE) = Learn with best explanations, minimal cost2025 Pro Tips:
- Start FREE: Gemini CLI + Codeium costs $0, surprisingly powerful!
- Best $20 value: ChatGPT Plus (includes Codex CLI + web chat)
- Best learning: Claude Code CLI (explains WHY, not just WHAT)
- Don’t confuse: Subscriptions ≠ API access (read CRITICAL section above!)
- Master ONE first: Don’t use 5 tools on day one, start simple!
Mental Model: AI as a Super-Intern
Section titled “Mental Model: AI as a Super-Intern”Here’s the best mental model for working with AI coding assistants:
AI is like a brilliant intern who:
- Knows syntax perfectly across 50+ languages
- Has read millions of code examples (all of GitHub!)
- Can write boilerplate instantly without typos
- Never complains about tedious tasks
- Works 24/7 without breaks
- Has photographic memory for APIs and documentation
- Doesn’t understand your business logic (unless you explain it clearly)
- Can confidently suggest wrong approaches (“hallucinations”)
- Needs you to verify their work (no critical thinking about correctness)
- Sometimes hallucinates nonexistent APIs or methods
- Doesn’t understand security implications
Layered analogy:
- Scale analogy: AI is to coding what calculators are to math - handles tedious computation, but you still need to know what to calculate
- Process analogy: AI is like a sous chef - preps ingredients (boilerplate), follows recipes (patterns), but you’re the head chef (architecture)
- Work analogy: AI is like autocorrect on steroids - catches obvious errors, suggests improvements, but can embarrassingly fail on context
Your role shifts from:
- Writing every line of code character by character
- Remembering every API method signature
- Fighting with syntax errors and typos
- Copy-pasting from Stack Overflow
To:
- Directing: Clearly explaining what you want (“specification as code”)
- Reviewing: Verifying AI output is correct, secure, efficient
- Refining: Iterating to get exactly what you need
- Integrating: Ensuring AI code fits your architecture and patterns
- Teaching: Explaining domain-specific context AI doesn’t have
Think of it as pair programming with an incredibly fast typist who needs guidance but executes brilliantly.
The AI Development Workflow
Section titled “The AI Development Workflow”Traditional Workflow (The Old Days)
Section titled “Traditional Workflow (The Old Days)”1. Think about problem (10 minutes)2. Search Google → Stack Overflow (20 minutes, 17 tabs open)3. Read documentation (30 minutes)4. Write code line by line (2-4 hours) ├─ Fight syntax errors (30 minutes) ├─ Look up API methods (20 minutes) └─ Refactor as you go (1 hour)5. Test (30 minutes)6. Debug failing tests (1-2 hours)7. Repeat steps 4-6 (3-4 iterations)8. Code review feedback (next day)9. Implement feedback (2 hours)
Total: 8-12 hours for a featureFrustration level: HighLearning: Moderate (lots of trial/error)AI-Augmented Workflow (The New Way)
Section titled “AI-Augmented Workflow (The New Way)”1. Think about problem (10 minutes)2. Write specification in natural language (15 minutes) "Create a Python function that..."3. AI generates code (30 seconds)4. Review and understand AI output (20 minutes) ├─ Ask AI to explain complex parts ├─ Request changes for edge cases └─ Verify security and correctness5. AI generates tests (30 seconds)6. Run tests, iterate with AI on failures (30 minutes)7. Integrate into codebase (20 minutes)8. AI generates documentation (30 seconds)
Total: 2-3 hours for same featureFrustration level: LowLearning: High (AI explains as it codes)Speedup: 3-5x faster for typical features
Time saved: 5-9 hours per feature
Extrapolated: 25-45 hours saved per week = 100-180 hours/month = 1-2 extra months of productivity per year!
The “AI-First” Mindset Shift
Section titled “The “AI-First” Mindset Shift”OLD mindset: “How do I code this?” NEW mindset: “How do I explain this to AI so it codes it correctly?”
This is a skill. Just like prompt engineering (Module 2!), directing AI to write code is a skill you’ll develop throughout this curriculum.
The best AI programmers:
- Write clear specifications in natural language (precision matters!)
- Break complex tasks into smaller, manageable steps
- Verify AI output carefully (trust, but verify)
- Know when AI is likely to struggle (novel problems, security)
- Iterate quickly when AI gets it wrong (refinement loop)
- Use AI to explain code they don’t understand (learning accelerator)
Shift from “coder” to “code reviewer + architect”
When to Use AI vs Traditional Coding
Section titled “When to Use AI vs Traditional Coding”AI Excels At (Use AI Confidently)
Section titled “AI Excels At (Use AI Confidently)”1. Boilerplate Code (AI saves hours)
Section titled “1. Boilerplate Code (AI saves hours)”You: "Create a FastAPI server with:- CORS enabled for localhost:3000- JSON logging- Health check endpoint- Error handling middleware- Request timing middleware"
AI: [Generates 150 lines of production-ready code in 10 seconds]
Traditional time: 1-2 hoursAI time: 2 minutes review + 5 minutes integrationSpeedup: 30-60x2. Pattern Repetition (AI is tireless)
Section titled “2. Pattern Repetition (AI is tireless)”You: "Add error handling with try/except to all 23 API endpoints- Catch specific exceptions- Log errors- Return appropriate HTTP status codes- Don't change existing logic"
AI: [Updates all endpoints consistently in 1 minute]
Traditional time: 2-3 hoursAI time: 5 minutes reviewSpeedup: 24-36x3. Code Translation (AI knows all languages)
Section titled “3. Code Translation (AI knows all languages)”You: "Convert this JavaScript React component to TypeScriptAdd proper types, interfaces, and generic constraints"
AI: [Converts perfectly, adds types you forgot]
Traditional time: 1-2 hours (learning TypeScript patterns)AI time: 10 minutes reviewSpeedup: 6-12x4. Documentation & Comments (AI never procrastinates)
Section titled “4. Documentation & Comments (AI never procrastinates)”You: "Add comprehensive docstrings to all functions in this moduleUse Google style, include examples, type hints"
AI: [Adds perfect documentation to 50 functions]
Traditional time: 3-4 hours (nobody likes writing docs)AI time: 20 minutes reviewSpeedup: 9-12x5. Debugging (AI sees patterns you miss)
Section titled “5. Debugging (AI sees patterns you miss)”You: "Why is this throwing 'list index out of range'?[pastes 50 lines of code]"
AI: "Line 23: You're iterating with range(len(items)+1)This causes index out of range on the last iteration.Change to: range(len(items))
Also, line 15 has a subtle bug where..."
Traditional time: 30 minutes to 2 hours (depending on complexity)AI time: 30 secondsSpeedup: 60-240x6. Refactoring (AI knows best practices)
Section titled “6. Refactoring (AI knows best practices)”You: "Refactor this 300-line function into smaller functions- Extract reusable logic- Follow Single Responsibility Principle- Preserve all functionality- Add type hints"
AI: [Refactors into 8 well-named functions, adds types, tests still pass]
Traditional time: 2-4 hoursAI time: 30 minutes review + testingSpeedup: 4-8x7. Learning (AI is a patient teacher)
Section titled “7. Learning (AI is a patient teacher)”You: "Explain how Python decorators work- Include examples- Show common patterns- Explain when to use them- Show me how to write my own"
AI: [Comprehensive tutorial with 5 examples, visual diagrams, best practices]
Traditional time: 1-2 hours reading docs, tutorials, Stack OverflowAI time: 10 minutes reading AI explanationSpeedup: 6-12xOverall: AI is a 10-100x multiplier for these tasks!
AI Struggles With (Use Traditional Approach or Heavy Verification)
Section titled “AI Struggles With (Use Traditional Approach or Heavy Verification)”1. Business Logic Unique to Your Domain
Section titled “1. Business Logic Unique to Your Domain”- AI doesn’t know your company’s specific rules
- Example: “Calculate commission based on our tier system” - AI doesn’t know your tiers
- Solution: Explain the rules clearly, or code it yourself then have AI generate tests
2. Novel Algorithms
Section titled “2. Novel Algorithms”- If it’s not in training data (published before 2024), AI will struggle
- Research papers, cutting-edge techniques
- Solution: Implement yourself, ask AI for optimization suggestions
3. Complex Architecture Decisions
Section titled “3. Complex Architecture Decisions”- “Should I use microservices or monolith?”
- “Which database best fits this use case?”
- AI can explain trade-offs but can’t know your constraints
- Solution: You decide, AI implements
4. Security-Critical Code
Section titled “4. Security-Critical Code”- Always verify AI-generated auth, crypto, validation
- AI can introduce subtle security bugs:
- SQL injection vulnerabilities
- Insecure password hashing
- Missing input validation
- Timing attacks
- Solution: Code yourself or HEAVILY review AI output with security lens
5. Performance-Critical Code
Section titled “5. Performance-Critical Code”- AI’s first solution may not be optimized
- Often generates O(n²) when O(n) is possible
- Solution: Profile and optimize yourself, ask AI for algorithmic improvements
6. Debugging Subtle Edge Cases
Section titled “6. Debugging Subtle Edge Cases”- AI can help, but complex bugs still need human intuition
- Especially: race conditions, memory leaks, distributed systems bugs
- Solution: Use AI as brainstorming partner, but you drive
7. Creative Problem-Solving
Section titled “7. Creative Problem-Solving”- Novel solutions to novel problems
- AI suggests conventional approaches first
- Solution: Brainstorm with AI, but innovate yourself
The AI Confidence Decision Matrix
Section titled “The AI Confidence Decision Matrix”| Task Type | AI Reliability | Your Role | Verification Level |
|---|---|---|---|
| Boilerplate | 95% | Quick review | Low (skim code) |
| Standard patterns | 90% | Verify logic | Medium (understand flow) |
| Code explanation | 90% | Learn & verify | Low (AI rarely lies about what code does) |
| Debugging common issues | 85% | Guide AI to fix | Medium (test the fix) |
| Test generation | 85% | Review coverage | Medium (run tests) |
| Refactoring | 75% | Review architecture | High (deep review) |
| New feature | 60% | Spec clearly, verify deeply | High (thorough testing) |
| Documentation | 95% | Quick review | Low (AI is great at docs) |
| Business logic | 40% | You drive, AI assists | Very High (you’re the expert) |
| Security code | 30% | You code, AI suggests | Critical (security review) |
| Novel algorithm | 20% | Research yourself | Very High (AI often wrong) |
Rule of thumb:
- If it’s been done a million times before (CRUD, APIs, tests), AI will excel
- If it’s unique to your situation, you’ll need to guide AI heavily or code it yourself
- If it’s security-critical, NEVER trust blindly
AI Coding Best Practices (Critical for Success!)
Section titled “AI Coding Best Practices (Critical for Success!)”These practices separate effective AI-augmented developers from those who struggle:
1. Always Specify Type Hints When Asking AI
Section titled “1. Always Specify Type Hints When Asking AI”Why: AI generates better code with type context, and you get better autocomplete
# BAD (vague request)You: "Create a function to process users"
AI: [Generates generic code with 'Any' types]
# GOOD (specific request)You: "Create a function that:- Takes: List[Dict[str, Any]] of users- Returns: List[User] (Pydantic model)- Filters users where age > 18- Add type hints and docstring"
AI: [Generates precise, type-safe code]Benefit: Better AI suggestions, fewer bugs, better tooling support
2. Request Tests Alongside Code
Section titled “2. Request Tests Alongside Code”Why: AI-generated code without tests is a ticking time bomb
# BAD (code only)You: "Create user validation function"
# GOOD (code + tests)You: "Create user validation function- Validates email, password strength, age- Include: - Implementation with type hints - Pytest tests covering edge cases - Test for empty inputs, invalid formats - Docstring with examples"
AI: [Generates function + comprehensive tests]Benefit: Catch AI mistakes immediately, ensure correctness, documentation
3. Iterate in Small Steps (Don’t Ask for Everything at Once)
Section titled “3. Iterate in Small Steps (Don’t Ask for Everything at Once)”Why: Large requests lead to AI errors, small iterations give you control
# BAD (too much at once)You: "Create complete user management system with CRUD, auth, role-based access, email verification, password reset, session management, and audit logging"
AI: [Generates 500 lines with subtle bugs, missing edge cases]
# GOOD (iterative approach)You: "Step 1: Create User model with Pydantic"AI: [Generates model]
You: "Step 2: Add CRUD operations for User"AI: [Adds CRUD, you review]
You: "Step 3: Add password hashing with bcrypt"AI: [Adds secure hashing, you verify]
[Continue iterating...]Benefit: Easier to review, catch mistakes early, maintain control
4. Always Review for Security Vulnerabilities
Section titled “4. Always Review for Security Vulnerabilities”Why: AI doesn’t think like an attacker
Common AI security mistakes:
# AI might generate (SQL injection risk!)query = f"SELECT * FROM users WHERE username='{username}'"
# You verify and fixquery = "SELECT * FROM users WHERE username = %s"cursor.execute(query, (username,))
# AI might generate (timing attack risk!)if user.password == provided_password: return True
# You fiximport secretsif secrets.compare_digest(user.password, provided_password): return TrueSecurity checklist:
- No SQL injection (use parameterized queries)
- No command injection (validate/sanitize input)
- No XSS (escape output)
- Secure password hashing (bcrypt/argon2, not sha256)
- No hardcoded secrets
- Input validation on all user input
5. Ask AI to Explain Unfamiliar Code
Section titled “5. Ask AI to Explain Unfamiliar Code”Why: Understanding > blind copy-paste
# AI generates code you don't fully understand
You: "Explain this code line-by-line:[paste the AI-generated code]
Particularly:- Why use `functools.lru_cache`?- What's the time complexity?- Are there edge cases I should test?"
AI: [Detailed explanation]
# Now you understand AND learned something!Benefit: Learn while building, avoid cargo-cult programming
6. Provide Examples of Your Coding Style
Section titled “6. Provide Examples of Your Coding Style”Why: AI matches patterns you show it
You: "I have this function in my codebase:
def process_payment( amount: Decimal, currency: str, user_id: UUID,) -> PaymentResult: '''Process payment with idempotency.''' try: # Implementation pass except PaymentError as e: logger.error(f'Payment failed: {e}', extra={'user_id': user_id}) raise
Write a similar function for refunds, matching this style:- Same type hints pattern- Same error handling- Same logging- Return RefundResult"
AI: [Generates code matching your style perfectly]Benefit: Consistent codebase, less refactoring, faster code review
7. Use AI for Code Review (Automated First Pass)
Section titled “7. Use AI for Code Review (Automated First Pass)”Why: Catch obvious issues before human review
You: "Review this code for:- Bugs and edge cases- Performance issues- Security vulnerabilities- Code smells- Better patterns
[paste your code]"
AI: [Detailed review with suggestions]
# Fix issues, THEN send to human reviewBenefit: Faster reviews, learn best practices, fewer embarrassing bugs
8. Version Control AI Changes (Git Best Practices)
Section titled “8. Version Control AI Changes (Git Best Practices)”Why: AI can make breaking changes, you need rollback capability
# BAD (let AI modify 10 files at once)git status# 10 files changed, hard to review
# GOOD (AI changes one thing at a time)git add feature.pygit commit -m "Add user validation (AI-generated)"
git add tests/test_feature.pygit commit -m "Add tests for user validation (AI-generated)"
# Easy to review, easy to rollbackBenefit: Granular control, easy rollback, clear history
Common Mistakes: Learn From Others’ Pain
Section titled “Common Mistakes: Learn From Others’ Pain”Mistake #1: “I Blindly Accepted AI’s First Suggestion”
Section titled “Mistake #1: “I Blindly Accepted AI’s First Suggestion””Symptom: Code looks good, passes basic tests, then explodes in production
Real Example:
# You asked: "Parse JSON from API response"
# AI generated:def parse_response(response: str) -> dict: return json.loads(response)
# You accepted without thinking# In production: API returns malformed JSON → JSONDecodeError → service crashWhy It’s Bad:
- No error handling
- Assumes API always returns valid JSON (it doesn’t!)
- One bad response takes down your service
The Fix:
# Always review and add error handlingfrom typing import Optional
def parse_response(response: str) -> Optional[dict]: try: return json.loads(response) except json.JSONDecodeError as e: logger.error(f"Failed to parse JSON: {e}", extra={"response": response}) return NonePrevention: NEVER ship AI code you haven’t reviewed and tested!
Mistake #2: “AI Generated 500 Lines, I Didn’t Understand It”
Section titled “Mistake #2: “AI Generated 500 Lines, I Didn’t Understand It””Symptom: Code works… until it doesn’t, and you have no idea why
Real Example:
# You asked: "Create async task queue with Redis"
# AI generated 500 lines of code with:# - Custom retry logic# - Connection pooling# - Distributed locks# - Dead letter queue
# You copy-pasted# 6 months later: Redis connection leak, nobody knows whyWhy It’s Bad:
- You can’t debug what you don’t understand
- You can’t modify it safely
- You can’t explain it to teammates
The Fix:
- Ask AI to explain complex parts: “Explain this retry logic”
- Request simpler version: “Simplify this, use existing library”
- Build incrementally: “Start with basic queue, add features one-by-one”
Rule: If you can’t explain the code to a teammate, you don’t understand it enough to ship it!
Mistake #3: “I Asked Vaguely, Got Garbage Code”
Section titled “Mistake #3: “I Asked Vaguely, Got Garbage Code””Symptom: AI generates something, but it’s not what you wanted
Real Example:
# Vague requestYou: "Make a function to filter data"
AI: [Generates generic filter that doesn't match your needs]
You: "No, not like that"AI: [Different wrong answer]
You: "Ugh, AI is useless!"Why It’s Bad: Garbage specification in = garbage code out
The Fix: Be extremely specific
# Specific requestYou: "Create a function that:
Input: List[Dict[str, Any]] where each dict has: - 'created_at': datetime - 'status': str (one of: 'pending', 'completed', 'failed') - 'amount': float
Output: List[Dict[str, Any]] filtered to: - created_at within last 7 days - status == 'completed' - amount > 100.0
Add: - Type hints - Docstring with example - Handle empty list gracefully
Name: filter_recent_high_value_transactions"
AI: [Generates exactly what you need]Rule: Spend 5 minutes writing a clear spec to save 30 minutes debugging vague code!
Mistake #4: “I Didn’t Test AI Code, Just Deployed It”
Section titled “Mistake #4: “I Didn’t Test AI Code, Just Deployed It””Symptom: Production bugs, customer complaints, late-night debugging
Real Example:
# AI generated email validation
# Looked good, you shipped it# In production: Rejects valid emails like "user+tag@example.com"# Customer support gets 50 complaintsWhy It’s Bad:
- AI doesn’t think of all edge cases
- AI training data might have bugs
- AI can hallucinate “facts” about how things work
The Fix: Always test, especially edge cases
# Request tests when you request codeYou: "Create email validator
Include pytest tests for:- Valid: standard email, with +, with subdomain- Invalid: no @, multiple @, invalid domain- Edge cases: unicode characters, very long emails"
AI: [Generates implementation + comprehensive tests]
# RUN THE TESTS before deployingRule: AI-generated code without tests is production-ready-to-fail!
AI Coding Patterns (Master These!)
Section titled “AI Coding Patterns (Master These!)”Pattern 1: The Specification Pattern
Section titled “Pattern 1: The Specification Pattern”Concept: Write detailed spec, AI generates perfect code
You: "Create a Python function that:
PURPOSE: Analyze text sentiment
INPUT:- text: str (1-10,000 characters)- language: str = "en" (ISO 639-1 code)
OUTPUT:- SentimentResult (Pydantic model) with: - score: float (range: -1.0 to 1.0, -1=very negative, 1=very positive) - label: Literal["negative", "neutral", "positive"] - confidence: float (range: 0.0 to 1.0)
BEHAVIOR:- Use transformers library, model: "cardiffnlp/twitter-roberta-base-sentiment"- If text length > 512 tokens, split and average scores- If language != "en", raise NotImplementedError- Handle empty text by returning neutral (score=0.0, confidence=0.0)
ERROR HANDLING:- Catch model loading errors- Catch tokenization errors
REQUIREMENTS:- Type hints- Comprehensive docstring with examples- Pytest tests (5+ test cases)
PERFORMANCE:- Cache model loading (load once per process)- Should process 100 texts/second on CPU"AI generates: Perfect implementation matching your exact specs
This is faster and more accurate than coding yourself!
Pattern 2: The Iteration Pattern
Section titled “Pattern 2: The Iteration Pattern”Concept: Refine through conversation, don’t expect perfection
# Iteration 1You: "Create function to validate email"AI: [Basic regex validation]
# Iteration 2You: "Add support for internationalized domains (unicode)"AI: [Updates with IDN support]
# Iteration 3You: "Add validation for common typos (gmial.com → gmail.com)"AI: [Adds typo detection and suggestions]
# Iteration 4You: "Add disposable email domain check"AI: [Integrates disposable email list]
# Result: Perfect email validator through 4 quick iterationsKey: Each iteration takes 30 seconds, total time: 5 minutes vs 2 hours coding from scratch
Pattern 3: The Example Pattern
Section titled “Pattern 3: The Example Pattern”Concept: Show AI your pattern, it replicates perfectly
You: "I have this function:
def get_user_by_id(user_id: UUID) -> Optional[User]: '''Fetch user by ID from database.''' try: return db.query(User).filter(User.id == user_id).first() except DatabaseError as e: logger.error(f'Failed to fetch user: {e}') return None
Write similar functions for:- get_user_by_email(email: str) -> Optional[User]- get_users_by_role(role: str) -> List[User]- get_active_users() -> List[User]
Match my style exactly."
AI: [Generates all 3 functions in your exact style]Benefit: Consistent codebase, AI learns your patterns
Pattern 4: The Explanation Pattern
Section titled “Pattern 4: The Explanation Pattern”Concept: Use AI as a teacher, not just code generator
You: "Explain this code line-by-line:
@lru_cache(maxsize=128)def fibonacci(n: int) -> int: return n if n < 2 else fibonacci(n-1) + fibonacci(n-2)
Questions:- What does @lru_cache do?- Why maxsize=128?- What's the time complexity with and without cache?- Are there better approaches?"
AI: "Detailed explanation:
Line 1: @lru_cache(maxsize=128) is a decorator that caches function results- First call fibonacci(5): computes and caches result- Second call fibonacci(5): returns cached result instantly- maxsize=128: keeps last 128 results in memory
Without cache:- Time complexity: O(2^n) - exponential, VERY slow- fibonacci(40) takes ~1 second
With cache:- Time complexity: O(n) - linear, fast!- fibonacci(40) takes 0.00001 seconds
Better approaches:1. Iterative (bottom-up): O(n) time, O(1) space2. Closed-form formula: O(1) but floating point errors
[Provides code examples for each]"
# You learned 3 new things while solving your problem!This is how you learn 10x faster with AI!
Pattern 5: The Debugging Pattern
Section titled “Pattern 5: The Debugging Pattern”Concept: AI as debugging assistant
You: "This code throws 'AttributeError: 'NoneType' object has no attribute 'name'':
def greet_user(user_id: int): user = get_user(user_id) return f'Hello, {user.name}!' # <-- Error here
Why? How to fix?"
AI: "The error occurs because get_user() returns None when user doesn't exist.
Root cause: Line 2, get_user(user_id) returns None for non-existent users
Fix options:
1. Defensive check:def greet_user(user_id: int) -> str: user = get_user(user_id) if user is None: return 'Hello, stranger!' return f'Hello, {user.name}!'
2. Better: Raise explicit error:def greet_user(user_id: int) -> str: user = get_user(user_id) if user is None: raise ValueError(f'User {user_id} not found') return f'Hello, {user.name}!'
3. Best: Type system catches this:def get_user(user_id: int) -> Optional[User]: # Makes None explicitdef greet_user(user_id: int) -> str: user = get_user(user_id) if user is None: # mypy would warn if you forgot this check raise ValueError(f'User {user_id} not found') return f'Hello, {user.name}!'
Recommendation: Use option 3 - type hints prevent this class of bugs!"AI debugs faster than Stack Overflow search, AND teaches you best practices!
Deep Dive (Optional): How AI Coding Assistants Actually Work
Section titled “Deep Dive (Optional): How AI Coding Assistants Actually Work”For the curious: What’s happening under the hood when AI writes code?
The Technology Stack
Section titled “The Technology Stack”1. Training Data:
- GitHub: 100+ million public repositories
- Stack Overflow: Questions and answers
- Documentation: Python docs, MDN, etc.
- Books: Programming textbooks
- Total: Trillions of tokens of code
2. Model Architecture:
Your prompt → Tokenizer → Transformer Model → Token Predictions → Detokenizer → Code output
Example:"Create fibonacci function" → [1234, 5678, 9012] (tokens) → gpt-5/Claude processes → [4567, 8901, 2345...] (output tokens) → "def fibonacci(n: int)..." (code)3. Context Window:
- Small models (Copilot): ~8K tokens (few files)
- Large models (Claude Code): 200K tokens (entire small codebase!)
One token ≈ 4 characters of code
So 200K tokens = 800,000 characters ≈ 10,000 lines of code!
Why AI Makes Mistakes
Section titled “Why AI Makes Mistakes”1. Training Data Bias:
- AI learns from public code (which contains bugs!)
- Old Stack Overflow answers (deprecated APIs)
- Popular but insecure patterns
2. Hallucination:
- AI “dreams up” APIs that don’t exist
- Example: AI might suggest
pandas.DataFrame.sort_by_column()(doesn’t exist, should besort_values())
3. Context Limits:
- Can’t see your entire monorepo if it’s > 200K tokens
- Might miss architectural constraints
4. No Execution:
- AI doesn’t RUN the code it generates (except ChatGPT Code Interpreter)
- Can’t verify it actually works
The “Temperature” Setting
Section titled “The “Temperature” Setting”Temperature controls randomness:
Temperature = 0.0: Deterministic (same input → same output)Temperature = 0.5: Balanced (default for most coding)Temperature = 1.0: Creative (different output each time)For coding: Use temperature = 0.2 for consistent, predictable code
Why low temperature?: Code has ONE correct answer (usually), you don’t want “creative” semicolon placement!
How Copilot vs Claude Code Differ
Section titled “How Copilot vs Claude Code Differ”Copilot:
Model: Codex (GPT-3.5 tuned on code)Context: Last ~50 lines of current fileSpeed: <500ms suggestionsApproach: Next-token prediction (autocomplete)Claude Code:
Model: Claude 3.5 SonnetContext: Entire project (up to 200K tokens)Speed: 2-10 seconds for responsesApproach: Reasoning + planning + executionAnalogy: Copilot is autocorrect, Claude Code is a research assistant
Did You Know? More Fascinating Facts
Section titled “Did You Know? More Fascinating Facts”The Origin of AI Coding Assistants
Section titled “The Origin of AI Coding Assistants”GitHub Copilot was released in June 2021 and was built on OpenAI Codex (GPT-3 fine-tuned on code from GitHub).
Initial reaction: “AI can’t really code! It just copy-pastes from training data!”
By 2023: 60% of developers using Copilot reported coding 50% faster
By 2024: AI coding assistants considered essential, like IDEs in the 1990s
Claude Code and Cursor emerged in 2024, pushing beyond autocomplete into agentic coding - AI that can reason about your entire project and take multi-step actions.
The transformation happened in ~3 years. That’s how fast AI development is moving!
AI Coding Stats (2024-2025)
Section titled “AI Coding Stats (2024-2025)”Productivity:
- 55% of professional developers use AI assistants daily
- 40% of code at companies using Copilot is AI-generated
- 3-5x productivity increase for boilerplate and tests
- 87% faster task completion for developers who use AI effectively
Time Allocation Shift:
- -60% time typing code
- +80% time reviewing code
- +40% time on architecture and design
Net result: You write MORE code in LESS time, but spend MORE time thinking and reviewing
What AI Can’t Do (Yet… 2025 Limitations)
Section titled “What AI Can’t Do (Yet… 2025 Limitations)”AI coding assistants in 2025 still can’t:
- Understand your product vision (you define requirements)
- Make architectural decisions (you choose patterns)
- Debug complex distributed systems (needs human intuition)
- Write perfectly secure code (always verify auth, crypto, validation)
- Replace code review (AI code still needs human review)
- Understand business context (unique to your company)
- Replace you (AI amplifies developers, doesn’t replace them)
The best developers in 2025: Know when to use AI, when to code themselves, and how to verify AI output
The Surprising Uses of AI Coding
Section titled “The Surprising Uses of AI Coding”Beyond writing code, developers use AI for:
- Learning new languages (30% faster onboarding)
- Understanding legacy code (“Explain this 1000-line function”)
- Writing documentation (AI loves this, humans hate it!)
- Code review comments (“Review this PR for issues”)
- Interview prep (“Ask me Python interview questions”)
- Pair programming (“I’m stuck, let’s brainstorm”)
- Rubber duck debugging (explaining to AI helps you find bugs!)
The rise of “AI pair programming” as default workflow
Try This: Interactive Challenges
Section titled “Try This: Interactive Challenges”Challenge 1: AI Tool Speed Test
Section titled “Challenge 1: AI Tool Speed Test”Goal: Compare autocomplete vs agentic AI speed
Task: Implement this function both ways:
def analyze_log_file(filepath: str) -> dict: """ Parse log file and return analytics: - total_lines: int - error_count: int - warning_count: int - most_common_errors: List[tuple[str, int]] """ passTry:
- With Copilot/autocomplete: Type function signature, let it autocomplete
- With Claude Code/agentic: Describe full spec, let it generate + tests
Time both approaches, compare!
Expected Results
Autocomplete approach:
- Time: 3-5 minutes (typing + corrections)
- Quality: Good basic implementation
- Tests: You write manually
Agentic approach:
- Time: 1-2 minutes (spec + review)
- Quality: Often better (handles edge cases)
- Tests: AI generates automatically
Lesson: Agentic AI is faster for complete features, autocomplete is faster for small tweaks
Challenge 2: The Debugging Race
Section titled “Challenge 2: The Debugging Race”Setup: I’ve intentionally broken this code with 5 bugs:
def calculate_discount(price: float, discount_percent: int) -> float: """Calculate final price after discount.""" discount = price * discount_percent / 100 final_price = price - discount return final_price
def apply_bulk_discount(prices: list[float]) -> list[float]: """Apply 10% discount to all prices.""" discounted = [] for i in range(len(prices) + 1): discounted.append(calculate_discount(prices[i], 10)) return discounted
# Testitems = [100, 200, 300]result = apply_bulk_discount(items)print(result) # Should print [90.0, 180.0, 270.0]Your Task:
- Find bugs manually (time yourself)
- Ask AI to find bugs (time that too)
Compare: Which was faster? Which found more bugs?
The 5 Bugs
-
Line 8:
range(len(prices) + 1)causes index out of range- Fix:
range(len(prices))
- Fix:
-
Type hint issue:
discount_percent: intbut should acceptfloattoo- Fix:
discount_percent: float
- Fix:
-
No input validation: What if price is negative?
- Fix: Add validation
-
No handling of empty list:
apply_bulk_discount([])works but no test- Fix: Add edge case handling
-
Potential float precision:
10.1 * 10 / 100might have rounding issues- Fix: Use
Decimalfor money calculations
- Fix: Use
AI will likely find all 5 in 10 seconds. Manual debugging: 3-10 minutes
Challenge 3: Specification Quality Experiment
Section titled “Challenge 3: Specification Quality Experiment”Hypothesis: Better specification → better AI output
Task: Ask AI to create a user registration function
Try 3 approaches:
1. Vague:
"Create user registration function"2. Moderate:
"Create function to register user- Takes email and password- Validates input- Saves to database- Returns success/failure"3. Detailed:
"Create user registration function:
INPUT:- email: str (must be valid email format)- password: str (min 8 chars, 1 uppercase, 1 number)- name: str (optional)
OUTPUT:- RegistrationResult (Pydantic): - success: bool - user_id: Optional[UUID] - errors: List[str]
BEHAVIOR:- Validate email format (RFC 5322)- Check password strength- Check if email already exists (db query)- Hash password with bcrypt (cost factor 12)- Save to User table- Send confirmation email (async)
ERROR HANDLING:- InvalidEmailError- WeakPasswordError- DuplicateEmailError
REQUIREMENTS:- Type hints- Comprehensive tests (10+ cases)- Docstring with examples"Compare: Quality, correctness, completeness of AI outputs
Expected Results
Vague → Generic code, missing validation, no error handling, no tests
Moderate → Decent code, basic validation, might miss edge cases
Detailed → Production-ready code, comprehensive tests, handles edge cases
Lesson: 5 minutes writing detailed spec saves 30 minutes fixing vague AI output!
️ Setting Up Your AI Coding Environment
Section titled “️ Setting Up Your AI Coding Environment”Option 1: VS Code + Claude Code (Recommended for Neural Dojo)
Section titled “Option 1: VS Code + Claude Code (Recommended for Neural Dojo)”Why This Setup:
- Claude Code can see your entire project (200K tokens)
- Explains reasoning clearly (best for learning!)
- Can run terminal commands (automated testing)
- Long context window (understands module context)
- Free VS Code + pay-per-use Claude API
Setup Steps:
- Install VS Code: https://code.visualstudio.com/
- Install Claude Code extension:
- Open VS Code
- Extensions (Cmd/Ctrl+Shift+X)
- Search “Claude Code”
- Install + Sign in with Anthropic account
- Configure API Key:
- Get key from: https://console.anthropic.com/
- Add to environment or let extension prompt you
- Test It:
- Open a Python file
- Press
Cmd/Ctrl+Shift+P→ “Claude Code: Start Chat” - Ask: “Explain this code”
Usage Tips:
Cmd/Ctrl+Shift+P→ “Claude Code: Start Chat” for new conversation- Select code → Right-click → “Ask Claude” for context-aware help
- Claude can read files, write code, run terminal commands
Cost: ~$3-5 for entire Neural Dojo curriculum (pay-per-use)
Option 2: GitHub Copilot (Great for Daily Coding)
Section titled “Option 2: GitHub Copilot (Great for Daily Coding)”Why This Setup:
- Fastest autocomplete (sub-second suggestions)
- Mature product, well-tested
- Great for writing functions, tests, boilerplate
- Works in VS Code, JetBrains, Vim, Neovim
Setup Steps:
- Sign up: https://github.com/features/copilot
- $10/month individual
- FREE for students (verify with GitHub Student Pack)
- Install Extension:
- VS Code: Search “GitHub Copilot” in extensions
- Install both “GitHub Copilot” and “GitHub Copilot Chat”
- Sign In: Extension prompts for GitHub login
- Test It:
- Create file:
test.py - Type:
def fibonacci( - Copilot suggests completion, press Tab to accept
- Create file:
Usage Tips:
- Write comment describing what you want, Copilot generates code
- Press
Alt+]/Alt+[to cycle through suggestions - Use Copilot Chat for explanations:
Ctrl+Shift+I
Cost: $10/month (FREE for students!)
Option 3: Cursor (All-in-One, Best Experience)
Section titled “Option 3: Cursor (All-in-One, Best Experience)”Why This Setup:
- Best of both worlds (autocomplete + agentic AI)
- Native AI integration, not an extension
- Composer mode for multi-file edits
- Import all your VS Code settings/extensions
Setup Steps:
- Download Cursor: https://cursor.sh/
- Import VS Code Settings:
- First launch → “Import from VS Code”
- Imports extensions, settings, keybindings
- Sign In: Free trial, then subscription
- Configure AI:
- Settings → AI → Choose model (Claude Sonnet recommended)
- Test It:
Cmd/Ctrl+K: Inline AI editingCmd/Ctrl+L: Chat with codebaseCmd/Ctrl+I: Composer (multi-file editing)
Usage Tips:
- Use autocomplete for small edits (like Copilot)
- Use Composer for complex refactoring (select files, describe change)
- Use Chat to understand codebase (“How does auth work?”)
Cost: $20/month (includes both autocomplete + agentic AI)
Recommended Setup for Neural Dojo
Section titled “Recommended Setup for Neural Dojo”Primary Setup (What we’ll use in examples):
VS Code + Claude CodeWhy:
- Free editor (VS Code)
- Pay only for what you use (~$3-5 total for curriculum)
- Best explanations (Claude is excellent at teaching)
- You can follow along with all examples
Optional Add-on (If Budget Allows):
+ GitHub Copilot ($10/month or FREE for students)Why:
- Fast autocomplete for daily coding
- Complements Claude Code (autocomplete for speed, Claude for complex tasks)
- Best experience: Both together!
Total Cost: $0-10/month depending on if you add Copilot
Module 1 Complete Checklist
Section titled “Module 1 Complete Checklist”Use this to verify you’re ready for Module 2:
Understanding
Section titled “Understanding”- I can explain the difference between autocomplete and agentic AI
- I understand when to use AI vs traditional coding
- I know the strengths and limitations of AI coding assistants
- I can name 3 AI coding best practices
- I have VS Code + Claude Code installed and working
- I’ve successfully asked Claude to generate code
- (Optional) I have Copilot or another autocomplete tool
- I can switch between chat and autocomplete modes
Practice
Section titled “Practice”- I’ve used AI to generate at least 100 lines of code
- I’ve reviewed AI output for bugs
- I’ve iterated on AI suggestions (not just accepted first output)
- I’ve asked AI to explain code I didn’t understand
Reflection
Section titled “Reflection”- I understand how AI will change my development workflow
- I know which tools I’ll use for which tasks
- I’m excited (or at least curious) about AI-driven development
All checked? You’re ready for Module 2: Prompt Engineering!
Further Reading
Section titled “Further Reading”Essential Articles
Section titled “Essential Articles”- “GitHub Copilot Research” - Academic study on Copilot effectiveness
- “The End of Programming” - Matt Welsh on AI’s impact
- “AI and the Future of Programming” - Industry analysis
Videos Worth Watching
Section titled “Videos Worth Watching”- Andrej Karpathy on AI Assistants - Insights from leading researcher
- GitHub Copilot Demo - Official introduction
- Cursor Tutorial - Official guide
Documentation & Tools
Section titled “Documentation & Tools”Autocomplete-First:
AI-First IDEs:
Terminal/CLI:
Agentic Extensions:
General AI for Coding:
️ Next Steps
Section titled “️ Next Steps”Congratulations! You now understand the AI development landscape and how to use AI as your coding partner.
You’ve learned:
- The 13 major AI coding tools (2025-verified) and when to use each
- How to think about AI as a “super-intern” coding partner
- When AI excels vs when to code yourself
- AI coding best practices (8 critical practices)
- Common mistakes and how to avoid them
- How to set up your AI coding environment
Next Module: Module 2: Prompt Engineering Fundamentals
In Module 2, you’ll learn:
- The art and science of prompt engineering
- How to structure prompts for best results
- Few-shot learning techniques
- Chain-of-thought prompting
- Prompt security and edge cases
- The Heureka Moment: Understanding that prompts are the new programming interface!
Why Module 2 is critical: AI coding assistants are only as good as your prompts. Master prompting, and you’ll be 10x more effective with every AI tool!
Ready? Let’s master prompt engineering in Module 2!
Last updated: 2025-11-22 (Fact-Checked for 2025 - All Pricing Verified) Module status: Complete + Fact-Checked Tools covered: 13 (Added: OpenAI Codex CLI, Gemini CLI) Critical addition: Subscriptions vs API Access section