Claude AI Fundamentals: A Developer's Reference Guide
Jan 18, 202615 min read
This guide covers the core concepts needed to work with Claude and AI-powered systems.
1. What is Claude
1.1 LLM (Large Language Model)
| Question | Answer |
|---|---|
| What | AI trained on massive text data that generates human-like responses by predicting the next word |
| Why | Enables natural language understanding, code generation, reasoning, and creative tasks |
| When | Use when you need text generation, analysis, coding assistance, or conversation |
Learn more: Mapping the Mind of a Language Model
1.2 Claude
| Question | Answer |
|---|---|
| What | Anthropic's LLM assistant (current family: Claude 4.5) |
| Why | Designed for helpfulness, harmlessness, and honesty |
| When | Use for coding, analysis, writing, reasoning tasks |
Models:
| Model | Best For |
|---|---|
| Opus 4.5 | Complex reasoning, coding, analysis |
| Sonnet 4.5 | Balanced speed/quality, daily tasks |
| Haiku 4.5 | Fast, simple tasks, high volume |
Learn more: Models Overview
2. How to Talk to Claude
Before learning prompt techniques, you need to know WHERE to interact with Claude.
2.1 Three Ways to Use Claude
| Method | Best For | Skill Level |
|---|---|---|
| claude.ai | Quick conversations, exploration | Beginner |
| API | Building apps, automation | Intermediate |
| Claude Code | Coding in terminal | Intermediate |
2.2 API (Application Programming Interface)
| Question | Answer |
|---|---|
| What | Code interface to send messages to Claude and receive responses |
| Why | Build apps, automate tasks, integrate Claude into systems |
| When | Building products, automation, custom integrations |
| Where | HTTP requests to api.anthropic.com |
Basic Flow:
Your Code → HTTP Request → Anthropic API → Claude → Response → Your Code
Learn more: API Overview
3. Core Concepts
Now that you know WHERE to use Claude, learn the concepts that affect HOW it responds.
3.1 Tokens
| Question | Answer |
|---|---|
| What | Units of text LLMs process (~4 characters or ~0.75 words per token) |
| Why | Determines cost (pay per token) and context limits (max tokens per conversation) |
| When | Consider when managing costs or hitting context limits |
Examples:
"Hello" = 1 token
"Hello, world!" = 4 tokens
1 page of text ≈ 500 tokens
Learn more: Token Counting
3.2 Context Window
| Question | Answer |
|---|---|
| What | Maximum tokens an LLM can "see" at once (Claude: up to 200K tokens, 1M in beta) |
| Why | Limits how much information Claude can consider in one conversation |
| When | Matters when working with long documents or extended conversations |
| Where | Applies to entire conversation: system prompt + messages + response |
Analogy: Context window is Claude's "working memory" - everything must fit in this space.
Learn more: Context Windows
3.3 Temperature
| Question | Answer |
|---|---|
| What | Controls randomness in responses (0 = deterministic, 1 = creative) |
| Why | Balances consistency vs. creativity based on task needs |
| When | Lower for factual tasks (code, math), higher for creative tasks (writing, brainstorming) |
| Where | Set as API parameter when making requests (see API Reference) |
Settings:
| Temperature | Use Case |
|---|---|
| 0.0 | Code generation, math, factual Q&A |
| 0.3-0.5 | General tasks, balanced responses |
| 0.7-1.0 | Creative writing, brainstorming |
3.4 System Prompt
| Question | Answer |
|---|---|
| What | Instructions that define Claude's behavior for a conversation |
| Why | Sets persona, rules, format, and constraints for all responses |
| When | Set once at conversation start, applies to all messages |
| Where | First message in API request, or CLAUDE.md file in Claude Code |
Example:
You are a senior TypeScript developer.
Always use strict typing.
Never use `any` type.
Explain your code changes briefly.
Learn more: System Prompts
4. Prompt Engineering
Now you know the concepts. Here's how to write effective prompts.
Reading List
| # | Topic | Link |
|---|---|---|
| 1 | Prompt Engineering Overview | Overview |
| 2 | Claude 4 Best Practices | Best Practices |
| 3 | Interactive Tutorial | GitHub Tutorial |
4.1 Be Explicit
| Question | Answer |
|---|---|
| What | State exactly what you need - leave nothing to interpretation |
| Why | Claude guesses when instructions are vague = inconsistent results |
| When | Always - especially for specific outputs |
| Vague (Bad) | Explicit (Good) |
|---|---|
| "Write a function" | "Write a TypeScript function that validates emails using regex, returns boolean" |
| "Make it better" | "Reduce complexity, extract duplicated logic, add null handling" |
4.2 Give Examples (Multishot Prompting)
| Question | Answer |
|---|---|
| What | Show 2-3 examples of input → output patterns |
| Why | Claude pattern-matches better than interpreting descriptions |
| When | Formatting, classification, data transformation |
Example:
Convert to slug format:
- "Hello World" → "hello-world"
- "My First Post!" → "my-first-post"
Now convert: "Claude AI Tutorial"
4.3 Structure with XML Tags
| Question | Answer |
|---|---|
| What | Use XML tags to create clear boundaries in prompts |
| Why | Long prompts become confusing - tags organize sections |
| When | Complex prompts with multiple parts (context, task, examples) |
Common Tags:
<context>Background info + WHY this matters</context>
<task>What you want Claude to do</task>
<examples>Input → Output patterns</examples>
<constraints>Rules and limitations</constraints>
<output_format>How to structure the response</output_format>
Learn more: Use XML Tags
4.4 Explain Why
| Question | Answer |
|---|---|
| What | Tell Claude why a rule matters, not just the rule |
| Why | Claude generalizes from reasons - applies to edge cases |
| When | Any constraint or rule in your prompt |
| Rule Only (Weak) | Rule + Why (Strong) |
|---|---|
| "Never use ellipses" | "Never use ellipses because TTS can't pronounce them" |
| "Keep responses short" | "Keep responses short - users read on mobile" |
4.5 Don't Say "Think"
| Question | Answer |
|---|---|
| What | The word "think" triggers extended thinking mode (slower, more tokens) |
| Why | Avoid unintended slow responses and higher costs |
| When | Everyday prompts where deep reasoning isn't needed |
| Avoid | Use Instead |
|---|---|
| "Think about..." | "Consider..." |
| "What do you think?" | "What's your assessment?" |
When TO use "think": Complex multi-step problems where you want deep reasoning.
4.6 Action Words vs. Discussion Words
| Question | Answer |
|---|---|
| What | Claude takes words literally - "suggest" = talk, "change" = edit |
| Why | Gets Claude to act instead of just discuss |
| When | You want Claude to make changes, not just recommend |
| Discussion (Claude talks) | Action (Claude does) |
|---|---|
| "Can you suggest improvements?" | "Improve this function" |
| "How would you fix this?" | "Fix this bug" |
4.7 Keep It Minimal
| Question | Answer |
|---|---|
| What | Explicitly tell Claude to avoid over-engineering |
| Why | Claude 4.5 loves to add extra files, abstractions, "future-proofing" |
| When | Coding tasks, implementation requests |
Add this to prompts:
Keep the solution minimal. Only change what's necessary.
Don't add abstractions or improvements beyond the request.
Three similar lines of code is better than a premature abstraction.
The Prompt Formula
YOUR PROMPT =
1. <context> Background info + WHY this matters
2. <task> EXPLICIT instruction with ACTION verbs
3. <examples> Input → Output (2-3 examples)
4. <constraints> Keep minimal. No over-engineering.
5. Tools & Agents
Claude can do more than generate text. With tools, it can interact with the real world.
Reading List
| # | Topic | Link |
|---|---|---|
| 1 | Tool Use Overview | Tool Use |
| 2 | Building Effective Agents | Building Agents |
| 3 | Agent SDK Overview | Agent SDK |
5.1 Tool Use
| Question | Answer |
|---|---|
| What | Functions Claude can request to execute (weather, database, APIs) |
| Why | Claude can only generate text - tools let it interact with the real world |
| When | Need Claude to fetch data, perform actions, access external systems |
| Where | Defined in API requests, executed on your system |
The Flow:
1. You → Claude: "What's the weather?" + tool definitions
2. Claude → You: "Call get_weather({location: 'Paris'})"
3. You → Claude: Execute tool, return "22°C, sunny"
4. Claude → You: "The weather in Paris is 22°C and sunny."
Tool Definition:
{
"name": "get_weather",
"description": "Get current weather. Use when user asks about weather.",
"input_schema": {
"type": "object",
"properties": {
"location": {"type": "string", "description": "City name"}
},
"required": ["location"]
}
}
Key Rule: Descriptions matter most - write 3-4 detailed sentences.
5.2 Agents
| Question | Answer |
|---|---|
| What | AI that uses tools autonomously in loops to complete complex tasks |
| Why | Handle multi-step tasks without human intervention at each step |
| When | Complex tasks requiring multiple actions (coding, research, data processing) |
| Where | Built on top of Claude API with tool use |
Agent = Claude + Tools + Loop
User Request
↓
Claude decides action
↓
Calls tool → Gets result
↓
Claude decides next action
↓
Calls another tool → Gets result
↓
... (repeats until done)
↓
Final response
5.3 Agentic Loop
| Question | Answer |
|---|---|
| What | Pattern: Claude calls tool → gets result → decides next action → repeats |
| Why | Enables autonomous multi-step problem solving |
| When | Tasks requiring exploration, iteration, or multiple actions |
| Where | Core pattern in Claude Code and custom agents |
Example: Code Fix Agent
1. Read file → sees error
2. Search codebase → finds related code
3. Edit file → makes fix
4. Run tests → sees failure
5. Edit file → adjusts fix
6. Run tests → passes
7. Done!
6. MCP & Claude Code
Advanced tools for power users and teams.
Reading List
| # | Topic | Link |
|---|---|---|
| 1 | MCP Overview | Model Context Protocol |
| 2 | MCP Servers | MCP Servers |
| 3 | MCP Course (Free) | MCP Course |
| 4 | Claude Code Overview | Claude Code Docs |
| 5 | Claude Code GitHub | GitHub Repo |
| 6 | Slash Commands | Slash Commands |
6.1 MCP (Model Context Protocol)
| Question | Answer |
|---|---|
| What | Standard protocol for connecting Claude to external tools and data |
| Why | One standard way to integrate any service (vs. custom integrations) |
| When | Connecting Claude to Slack, Jira, databases, APIs |
| Where | Between Claude Code and external services |
Analogy: MCP is like USB-C for AI - one standard connector for everything.
Claude Code ←→ MCP Protocol ←→ MCP Server ←→ External Service
(Slack, Jira, DB)
6.2 MCP Server
| Question | Answer |
|---|---|
| What | Service that exposes tools to Claude via MCP protocol |
| Why | Packages integration logic in reusable, standard format |
| When | Need Claude to interact with a specific service |
| Where | Runs locally or remotely, connects to Claude Code |
Examples:
slack-mcp-server → Post messages, read channels
jira-mcp-server → Create issues, query tickets
github-mcp-server → Create PRs, read code
6.3 Claude Code
| Question | Answer |
|---|---|
| What | Anthropic's CLI tool for coding with Claude in your terminal |
| Why | AI coding assistant that understands your entire codebase |
| When | Daily development - coding, debugging, git, documentation |
| Where | Runs in terminal, works in any project directory |
Key Features:
| Feature | What it does |
|---|---|
| Codebase understanding | Reads and searches your entire project |
| Multi-file editing | Makes changes across multiple files |
| Git integration | Commits, PRs, branch management |
| MCP integration | Connects to external tools |
6.4 CLAUDE.md
| Question | Answer |
|---|---|
| What | Project instructions file Claude reads automatically |
| Why | Gives Claude context about your project without repeating yourself |
| When | Set up once per project, Claude reads it every session |
| Where | Root of your project directory |
Contains:
- Project architecture overview
- Coding standards and conventions
- Build commands
- Important patterns to follow
6.5 Slash Commands
| Question | Answer |
|---|---|
| What | Custom shortcuts for complex prompts (e.g., /commit, /plan) |
| Why | Reusable workflows without retyping long prompts |
| When | Repetitive tasks with consistent patterns |
| Where | Defined in .claude/commands/ folder |
Examples:
/commit # Smart commit with good message
/review-pr 123 # Review pull request #123
/plan # Create implementation plan
Quick Glossary
| Term | One-liner |
|---|---|
| LLM | AI that predicts next words based on training data |
| Token | ~4 characters or ~0.75 words |
| Context Window | LLM's "memory" for current conversation |
| Temperature | Creativity dial (0=focused, 1=random) |
| System Prompt | Instructions defining Claude's behavior |
| API | Code interface to talk to Claude |
| Tool | Function Claude can request to execute |
| Agent | Claude + tools + loop = autonomous task completion |
| Agentic Loop | Tool call → result → next decision → repeat |
| MCP | USB-C for AI - standard protocol to connect tools |
| MCP Server | Service exposing tools via MCP protocol |
| Claude Code | Terminal-based AI coding assistant |
| CLAUDE.md | Project instructions file for Claude |
| Slash Command | Shortcut for complex prompts |
Share: