What is Claude?
Claude is a family of large language models built by Anthropic with a focus on safety, helpfulness, and honesty. For most users, the fastest way to start is directly inside your IDE — no setup required beyond installing the extension. This module covers what Claude is, how it was designed, what makes it different from other AI assistants, and what it can do for your team.
Who Built Claude — and Why It Matters
Anthropic was founded in 2021 by former members of OpenAI, including Dario Amodei and Daniela Amodei. The company's founding premise is that AI systems are going to be enormously powerful, and the organisations building them have a responsibility to make them safe by design, not as an afterthought. This is not marketing — it shapes every decision from training methodology to model release strategy.
Where other labs optimise primarily for benchmark scores or viral demos, Anthropic invests heavily in alignment research: the science of ensuring AI systems do what humans actually want them to do, even in edge cases. Claude is the product of that research. It is not a research project itself — it is a production-grade assistant — but its safety properties come from deep technical work, not just content filters bolted on at the end.
Safety-First Design
Anthropic builds Claude using Constitutional AI (CAI) — a technique where the model is trained against a written set of principles (a "constitution") rather than relying solely on human raters. This makes Claude's behaviour more consistent and auditable. The model is trained to be helpful, harmless, and honest — and these constraints hold across all interfaces, including Claude Code in VS Code.
Works Where You Work
Claude is available via Claude.ai in the browser, the Claude Code extension in VS Code and JetBrains, the desktop app, the mobile app, and the Anthropic API. This track focuses on the IDE experience because that is where most professional work happens — but the underlying model is the same everywhere.
200K Token Context
All Claude 4 models support up to 200,000 tokens — roughly 150,000 words. You can feed entire codebases, documents, or email threads into a single conversation. This is not a theoretical number: Claude maintains strong recall and reasoning quality across the full window, which is critical for real-world use cases like codebase analysis or long-document summarisation.
Agentic by Default
Claude Code can read files, write code, run terminal commands, search the web, and complete multi-step tasks — all from a chat panel inside VS Code, no Python script required. This "agentic" capability means Claude does not just answer questions — it takes actions, checks results, and iterates.
Constitutional AI — How Claude Learns to Be Helpful and Safe
Most language models are fine-tuned using Reinforcement Learning from Human Feedback (RLHF): humans rate model outputs, and the model learns to produce outputs that get higher ratings. This works, but it has a limitation — the model's behaviour is only as consistent as the pool of human raters, and it is hard to audit why the model behaves a certain way.
Anthropic developed Constitutional AI (CAI) as an extension to RLHF. The key idea:
Write a Constitution
Anthropic defines a set of written principles — the "constitution" — that describe how Claude should behave. For example: "Choose the response that is most helpful to the human while being honest and avoiding harm." These principles are explicit and auditable.
Self-Critique
Claude generates responses, then critiques its own responses against the constitution. "Does this response violate any of my principles? How could I improve it?" This self-evaluation step produces training signal without requiring human raters for every example.
Reinforcement Learning
The model is trained to prefer responses that pass its own constitutional review. Over many iterations, this produces a model that is reliably helpful while avoiding harmful, deceptive, or manipulative outputs — even in novel situations the human raters never saw.
CAI means Claude's safety behaviour is consistent and predictable. You will not get wildly different safety responses depending on phrasing. For business use — especially customer-facing applications — this consistency is critical. You can trust that Claude will decline harmful requests reliably, without being so cautious that it refuses legitimate work.
How Claude Differs from GPT, Gemini, and Other Models
All major language models share the same transformer architecture at their core. The differences are in training methodology, safety approach, product philosophy, and specific capabilities. Here is an honest comparison:
| Dimension | Claude (Anthropic) | GPT (OpenAI) | Gemini (Google) |
|---|---|---|---|
| Safety approach | Constitutional AI — principles-based, self-critiquing | RLHF with content moderation layers | RLHF with Google's safety filters |
| Context window | 200K tokens (all models) | 128K (GPT-4o), 1M (o3 via API) | Up to 2M tokens (Gemini 2.5) |
| Strengths | Long-form writing, nuanced reasoning, code generation, following complex instructions precisely | Broad general knowledge, large ecosystem, image generation | Multimodal (native audio/video), Google integration, very large context |
| IDE integration | Claude Code (VS Code, JetBrains) — deep agentic capability | GitHub Copilot — inline completion focused | Gemini Code Assist — Google Cloud focused |
| Transparency | Publishes alignment research, model cards, system prompt visible | Publishes some research, system prompt hidden by default | Limited alignment publications |
| Pricing model | Per-token (input/output priced separately) | Per-token or subscription | Per-token or included with Google Workspace |
Each model family has genuine strengths. Claude excels at careful instruction-following, long documents, and code. GPT has a massive ecosystem and strong image generation. Gemini has the largest context window and native Google integration. The right choice depends on your specific use case, existing infrastructure, and priorities around safety and transparency.
What Claude Can Do — A Capabilities Overview
Claude is a general-purpose assistant, but certain capabilities are particularly strong. Here is what you can expect across the most common use cases:
Writing & Content
Drafting emails, marketing copy, product descriptions, documentation, blog posts, press releases. Claude follows brand voice instructions precisely and handles negative constraints ("no exclamation marks") better than most models.
Code & Engineering
Writing, reviewing, debugging, and refactoring code across 20+ languages. Claude Code in VS Code can edit files, run tests, and fix errors autonomously. Particularly strong in Python, TypeScript, and Rust.
Analysis & Research
Summarising documents, extracting structured data from unstructured text, comparing options across criteria, identifying patterns in data. The 200K context window means you can feed entire reports without chunking.
Reasoning & Problem-Solving
Multi-step logical reasoning, mathematical analysis, strategic planning, root cause analysis. Claude Opus with extended thinking mode excels at problems requiring 10+ reasoning steps.
| Use Case | Example | Recommended Model |
|---|---|---|
| Quick classification or labeling | Is this email a complaint, a question, or praise? | Haiku |
| Email drafting | Draft a reply to a customer asking about return policy | Sonnet |
| Code generation | Write a REST endpoint for order lookup | Sonnet |
| Complex architecture review | Review our microservices design for single points of failure | Opus |
| Long document analysis | Summarise this 80-page contract and flag unusual clauses | Sonnet or Opus |
| Data extraction from images | Extract the table from this screenshot of a spreadsheet | Sonnet |
| Creative writing | Write a brand story for our About page | Opus |
What Claude Cannot Do — Honest Limitations
Understanding Claude's limitations is as important as understanding its capabilities. Here are the genuine constraints:
No Real-Time Information
Claude's training data has a knowledge cutoff. It does not know today's stock prices, yesterday's news, or your company's latest sales numbers unless you provide them. Claude Code can search the web, but base Claude cannot browse the internet on its own.
No Image Generation
Claude can analyse images but cannot create them. If you need image generation, you will need a separate tool like DALL-E, Midjourney, or Stable Diffusion.
Hallucination Risk
Like all language models, Claude can generate plausible-sounding but incorrect information — especially for obscure facts, precise numbers, or recent events. Always verify claims that matter. Claude is generally good at saying "I'm not sure" but it is not perfect.
No Persistent Memory
Each conversation starts fresh. Claude does not remember previous sessions unless you provide context via CLAUDE.md or attached files. This is a privacy feature, but it means you need to manage context deliberately.
The most dangerous limitation of any LLM is not that it makes mistakes — it is that it makes mistakes with confidence. Claude is trained to express uncertainty, but it does not always succeed. For any output that will be used in production, customer-facing communication, or decision-making, always have a human review step. This is not a Claude-specific issue — it applies to every language model on the market.
Getting Started: Claude Code in VS Code
Install the Extension
Open VS Code → Extensions (Ctrl+Shift+X) → search Claude Code → Install. Sign in with your Anthropic or Claude.ai account. No API key needed for claude.ai subscribers.
Open the Chat Panel
Click the Claude icon in the Activity Bar (left sidebar) or press Ctrl+Shift+C. The chat panel opens on the right. Type your first message — Claude can already see your open workspace.
Add Context with @
Type @filename to attach a specific file, or @folder to add a whole directory. Claude reads the content and can answer questions, suggest edits, or generate new code against it.
Apply Edits Inline
When Claude suggests a code change, a diff view appears. Click Accept to apply it directly to your file — no copy-pasting. Click Reject to discard.
Run Terminal Commands
Ask Claude to run commands: "run the tests" or "install the dependencies". Claude opens a terminal, executes the command, reads the output, and reports back — or fixes errors automatically.
ThreadCo's First Week with Claude
Here is how ThreadCo — a 3-person sustainable T-shirt brand — used Claude in their first week:
| Day | Task | What Claude Did | Time Saved |
|---|---|---|---|
| Monday | Customer email backlog (50 emails) | Drafted template replies for the 5 most common question types | 3 hours |
| Tuesday | Product description refresh | Rewrote 20 product descriptions in ThreadCo's brand voice | 4 hours |
| Wednesday | Bug in checkout flow | Identified null reference in webhook handler, proposed fix with test | 2 hours |
| Thursday | Competitor analysis | Analysed 3 competitor websites and produced a comparison matrix | 5 hours |
| Friday | Weekly report | Summarised 200 customer reviews into a 1-page executive summary | 4 hours |
When to Use Claude — and When Not To
Claude is powerful, but it is not the right tool for every task. Knowing when to reach for Claude and when to use a different approach is a skill in itself.
Use Claude When...
You need to draft, summarise, analyse, code, review, or explain. Tasks involving language comprehension and generation are Claude's sweet spot. Also use Claude when you need a "first draft" quickly — even if you will edit it heavily, starting from something is faster than starting from nothing.
Don't Use Claude When...
You need guaranteed mathematical precision (use a calculator or spreadsheet), real-time data (use APIs and databases directly), image generation (use DALL-E/Midjourney), or regulated outputs that require certification (legal filings, medical diagnoses). Claude can assist with all of these, but it should not be the sole source.
Augment, Don't Replace
The best workflows use Claude to augment human capability, not replace human judgment. Claude drafts, the human reviews. Claude identifies patterns, the human decides what to do about them. Claude generates code, the human tests and deploys it. This human-in-the-loop approach captures 80% of the speed benefit with near-zero risk.
The "5-Minute Rule"
If a task would take you less than 5 minutes to do yourself, just do it. The overhead of writing a good prompt, reviewing the output, and correcting any issues often exceeds 5 minutes for very simple tasks. Claude's value is highest on 30-minute to 4-hour tasks where it can compress the time to 5-15 minutes.
Anthropic's Approach to AI Safety
Understanding Anthropic's safety philosophy helps you predict how Claude will behave in edge cases and why it sometimes declines requests.
Responsible Scaling Policy
Anthropic has published a Responsible Scaling Policy (RSP) that defines capability thresholds. Before releasing more powerful models, Anthropic assesses risks and implements safeguards proportional to the model's capabilities. This is why Claude releases are sometimes slower than competitors — safety evaluation takes time.
Model Cards and Transparency
Anthropic publishes model cards for each Claude release, documenting capabilities, limitations, and known failure modes. They also share research on alignment, interpretability, and safety. This transparency helps users make informed decisions about where to deploy Claude.
Declining Harmful Requests
Claude will decline requests that could cause harm — generating malware, creating deceptive content, or assisting with illegal activities. Unlike a content filter that blocks keywords, Claude's safety training is nuanced: it can discuss sensitive topics in educational contexts while declining to provide actionable harmful instructions.
Data Privacy
Anthropic's enterprise plans offer zero data retention — your conversations are not stored or used for training. Consumer plans have different terms. For business use, verify your plan's data handling policies. Claude does not learn from individual conversations — it does not "remember" your data for future sessions or other users.
In day-to-day use, Claude's safety training is nearly invisible. It helps you with legitimate tasks without friction. The only time you will notice it is if you ask for something that could be genuinely harmful. If Claude declines a legitimate request, rephrase it with clear professional context: "I need this for [specific business purpose]." Claude is designed to be helpful to well-intentioned users.
How to Access Claude — All Channels
| Channel | Best For | Key Feature |
|---|---|---|
| Claude.ai (browser) | General conversations, document analysis, quick questions | No setup required, supports file uploads and projects |
| Claude Code (VS Code / JetBrains) | Software development, code review, multi-file editing | Deep IDE integration, file system access, terminal control |
| Claude Desktop App | Daily assistant work, MCP tool connections | Native app, supports MCP servers for external data |
| Claude Mobile App | On-the-go questions, voice conversations | Voice input, camera for image analysis |
| Anthropic API | Custom integrations, automated workflows, production systems | Full programmatic control, streaming, tool use, batching |
| Amazon Bedrock / Google Vertex | Enterprise deployments with existing cloud contracts | Use Claude through your existing AWS or GCP billing |
If your team needs to integrate Claude into a custom app or automate calls programmatically, see Module 19: API Usage. This track (Modules 07–24) focuses on day-to-day IDE use — no coding required.
Pricing Overview
Claude is available through several pricing models depending on how you access it:
| Plan | Price | What You Get | Best For |
|---|---|---|---|
| Free | $0/month | Limited daily messages with Sonnet, basic file uploads | Trying Claude for the first time |
| Pro | $20/month | Higher message limits, Opus access, Projects feature, Claude Code in VS Code | Individual professionals, small teams |
| Team | $30/user/month | Everything in Pro + admin controls, shared projects, higher limits | Teams of 5+ who need shared workflows |
| Enterprise | Custom pricing | Everything in Team + SSO, zero data retention, dedicated support, custom terms | Large organisations with compliance requirements |
| API | Per-token | Programmatic access, full control, all models, all features | Developers building custom integrations |
Claude Code in VS Code is included with Pro, Team, and Enterprise plans at no additional cost. You do not need an API key to use Claude Code — your claude.ai subscription covers it. API usage (for custom integrations) is billed separately per token.
Claude in the AI Landscape — 2026 Context
The AI assistant market is evolving rapidly. Here is where Claude fits in the broader landscape as of April 2026:
Frontier Model Competition
Claude Opus 4.6, GPT o3, and Gemini 2.5 Pro are the current frontier models. They are broadly competitive on most benchmarks, with each having specific strengths. The days of one model being "clearly best at everything" are over — model selection is now about fit for purpose.
Open Source Alternatives
Models like Llama 3.3 and Mistral Large offer strong capabilities that you can self-host. These are viable for privacy-sensitive use cases or high-volume low-complexity tasks. However, they require infrastructure expertise and generally trail frontier models on complex reasoning.
Specialised Tools
For specific tasks — image generation (Midjourney, DALL-E), voice (ElevenLabs), search (Perplexity), music (Suno) — specialised tools often outperform general-purpose models. Claude excels as a general reasoning and language engine; complement it with specialised tools for specific media types.
The Integration Layer
Claude's MCP protocol, tool use capabilities, and agentic behaviour position it as a coordination layer that connects to specialised tools. Rather than doing everything itself, Claude orchestrates: calling databases, triggering APIs, coordinating workflows, and synthesising results from multiple sources.
Key Terminology
| Term | Definition |
|---|---|
| Token | The basic unit Claude reads and writes. One token is roughly 3/4 of a word. "ThreadCo" is 2 tokens. A 200K token context window holds roughly 150,000 words. |
| Context window | The total amount of text Claude can consider in a single conversation — your messages, Claude's responses, system instructions, and attached files all share this space. |
| System prompt | Hidden instructions that set Claude's behaviour for a conversation. In Claude Code, your CLAUDE.md file serves this role. In the API, you set it explicitly. |
| Temperature | A setting that controls randomness. Lower temperature (0.0) = more deterministic, predictable outputs. Higher temperature (1.0) = more creative, varied outputs. |
| Agentic | Claude acting autonomously across multiple steps — reading files, running commands, checking results, and iterating — rather than just answering a single question. |
| MCP (Model Context Protocol) | An open protocol that lets Claude connect to external tools and data sources — databases, APIs, CRMs — through a standardised interface. |
| CLAUDE.md | A markdown file at the root of your project that Claude Code reads automatically at the start of every session. Use it for standing instructions, project context, and conventions. |
Hands-On Exercises
Open Claude Code in VS Code. Without attaching any files, ask: "What can you see in my workspace right now?" Note what Claude reports. Then attach a file with @ and ask the same question. Compare the two responses. This teaches you the difference between implicit and explicit context.
Create a short CLAUDE.md file with 3 brand voice rules (e.g., "Always be friendly", "Never use exclamation marks", "Mention sustainability in every product description"). Ask Claude to write a product description. Check whether all 3 rules were followed. Adjust and re-test.
Take a moderately complex task (e.g., "Explain the pros and cons of microservices vs monolith for a 3-person team"). Run it with Haiku, then Sonnet, then Opus. Compare the depth, nuance, and accuracy of each response. Document when the cheaper model was "good enough" and when you genuinely needed the more capable one.
Ask Claude a question about something that happened last week (a recent news event, a sports score, etc.). Observe how Claude handles the knowledge cutoff. Then ask Claude to search the web for the answer (in Claude Code). Compare the two approaches and note when web search is necessary.
Ask Claude Code to run a simple terminal command in your workspace (e.g., ls or git status). Then ask it to run something that will produce an error (e.g., npm test in a directory with no tests). Observe how Claude reads the output, interprets the error, and suggests a fix. This demonstrates the agentic loop.
Key Takeaways
Claude is Built Differently
Constitutional AI, safety-first design, and transparent alignment research make Claude predictable and trustworthy for professional use. You get consistent, helpful behaviour without having to "trick" the model into being useful.
Start in Your IDE
The fastest path to productivity is Claude Code in VS Code. No API keys, no setup, no code. Install the extension, open the chat panel, and start asking questions about your project. Everything else builds on this foundation.
Context is Everything
Claude is only as good as the context you give it. Attach relevant files with @, create a CLAUDE.md for standing instructions, and be specific about what you want. The difference between generic output and brilliant output is almost always about context.
Human + AI > Either Alone
The best workflow is not "let Claude do it all" or "do it all yourself." It is: Claude drafts, you review. Claude identifies, you decide. Claude generates, you verify. This partnership captures the speed of AI with the judgment of a human.
Maker: Anthropic (founded 2021, San Francisco). Training approach: Constitutional AI (CAI). Models: Opus (most capable), Sonnet (balanced), Haiku (fastest). Context window: 200K tokens (all models). Access: Claude.ai, Claude Code (VS Code/JetBrains), Desktop, Mobile, API. Key differentiator: Safety-first design, strong instruction following, deep IDE integration. Limitations: No image generation, no persistent memory, knowledge cutoff, potential for hallucination.
Most teams see immediate value from Claude within the first day — quick questions, draft generation, and code explanations work out of the box. The real productivity gains come in weeks 2-4, as you learn to write better prompts, build CLAUDE.md files, and develop team conventions. The modules ahead are designed to accelerate this learning curve. Treat the exercises seriously — they are the fastest path from "this is useful" to "this is indispensable."