Most people know Claude as a chatbot. But Anthropic now ships three distinct products, and they do very different things. We use all three daily at RIVER. Here's a practical guide to which one fits how you work.
The Short Version
- Claude is your thinking partner. You chat, it responds. Best for research, writing, analysis, and reasoning through problems.
- Claude Cowork is your knowledge worker. You describe an outcome, step away, and come back to finished work. Best for complex tasks across files, spreadsheets, presentations, and research synthesis.
- Claude Code is the most powerful of the three. It lives in your terminal, reads your entire codebase, writes code, runs tests, and manages git. If you want to excel in the AI era, this is the one to learn.
These tools are improving fast. Anthropic's latest model release (Opus 4.7, April 2026) resolves 3x more production coding tasks than its predecessor, scores 70% on CursorBench, and shows a 14% lift on complex multi-step workflows. The gap between "AI assistant" and "AI worker" is closing every quarter. Picking the right product matters more than ever.
Claude: Chat. Think. Draft.
Claude is the product most people start with. It's a conversational AI you interact with at claude.ai or on your phone. You ask questions, upload documents, and get thoughtful responses.
What makes Claude more than a chatbot is Projects. Create a project space, upload your key reference documents (a strategy doc, a brand guide, your last board report), add a few lines of custom instructions, and Claude remembers that context across every conversation in that project.
Who it's for: Anyone who does knowledge work. CEOs drafting strategy. Analysts reviewing reports. Marketers refining copy. Lawyers summarising contracts. If your job involves thinking, writing, or making sense of information, Claude is where you start.
What it does well:
- Deep reasoning through complex problems
- Drafting and refining documents with context
- Analysing uploaded files (PDFs, spreadsheets, images)
- Explaining technical concepts in plain language
- Extended thinking that shows its reasoning process
Limitations:
- Context windows cap how much it can hold at once. Upload 50 documents and it won't recall all of them equally. Start with 3-5 key files per project and expand from there.
- It can be confidently wrong. Claude sounds certain even when it's guessing. Always verify critical facts, numbers, and citations before acting on them.
- Projects are the unlock. Without them, every conversation starts from scratch and you lose continuity between sessions.
- No file access. Claude can analyse what you upload, but it can't read your local files, create spreadsheets, or organise folders. That's Cowork.
Getting started (5 minutes):
- Go to claude.ai and create an account
- Create a new Project
- Upload 3-5 key documents you reference regularly
- Add 2 lines of custom instructions: who you are and what you need
- Start a conversation. You now have a contextual assistant that knows your world.
I use Claude Projects for every major initiative. Upload the brief, the constraints, the stakeholder feedback. Then I can ask "given everything you know about this project, what am I missing?" and get an answer that actually accounts for the full picture.
Tim Hatherley-Greene
Chief Operating Officer
Claude Cowork: Describe. Delegate. Done.
This is the one most people haven't tried yet, and it's the biggest shift in how AI tools work.
Claude Cowork turns Claude from a conversation partner into an agent that does work on your behalf. You don't chat back and forth. You describe the outcome you want, Claude makes a plan, and then it executes. You come back to finished work.
It runs in Claude Desktop (macOS and Windows) and has direct access to your local files. That means it can read your documents, create new ones, organise folders, and build real outputs like Excel spreadsheets with working formulas, formatted PowerPoint presentations, and structured reports.
Who it's for: Knowledge workers who spend hours on tasks that could be described in one sentence. Operations leads, project managers, analysts, executives who need outputs, not conversations.
What it does well:
- Creating spreadsheets with formulas, charts, and conditional formatting
- Organising files and folders by type, date, or project
- Synthesising research from multiple documents into structured summaries
- Processing receipts, invoices, or data into organised reports
- Running tasks on a schedule (daily summaries, weekly data pulls)
- Working while you step away. Describe the task, go to lunch, come back to results.
What sets it apart from Claude chat:
- It takes action on your files, not just talks about them
- It breaks complex work into parallel subtasks automatically
- It creates real file outputs (not just text in a chat window)
- It can run for extended periods without timing out
- You can assign tasks from your phone while your desktop does the work
Limitations:
- Your desktop app has to stay open. Close it and active tasks stop. You can't start something, shut your laptop, and expect it to finish.
- Complex multi-step tasks sometimes drift. It'll get 80% right and miss a nuance on the last 20%. Check in on longer tasks rather than assuming the output is perfect.
- Not suitable for regulated workloads. Cowork activity isn't captured in audit logs or compliance exports. Don't use it for anything you need to prove to an auditor.
- It burns through your usage allocation faster than chat. Each task uses multiple model calls behind the scenes. Batch related work together and use regular Claude for simpler questions.
- It's still early. Cowork is the newest of the three products and it shows. The capabilities are real, but the edges are rougher than Claude chat.
Getting started (5 minutes):
- Download Claude Desktop
- Switch to the Cowork tab (you'll see "Chat" and "Cowork" at the top)
- Try: "Organise my Downloads folder by file type, then by month"
- Watch it plan the work, execute, and deliver
Where Cowork Shines
Cowork is at its best when the task would take you 30-60 minutes of repetitive work: compiling data from multiple files, reformatting a report, processing a folder of documents. If you can describe the outcome in one sentence but it would take an hour to do manually, that's a Cowork task.
Claude Code: The One to Learn
We saved the strongest for last.
Claude Code is the most capable of the three products, and the one we'd recommend to anyone serious about working in the AI era. It's not just for developers anymore.
It runs in your terminal, in VS Code, in JetBrains IDEs, and in the Claude Desktop app. It reads your entire codebase (or document set), understands project structure and conventions, writes and edits across multiple files, runs commands, manages git, and executes shell commands. The full capability list is long and growing fast.
This isn't autocomplete. It's an agentic system that plans, executes, and verifies.
Who it's for: Developers, yes. But also technical leads, data analysts, and anyone willing to work in a terminal. If you can type a command, you can use Claude Code. And we think more people should.
What it does well:
- Understanding unfamiliar codebases in minutes (not hours)
- Building features across multiple files with consistent patterns
- Writing and running tests, fixing what fails
- Refactoring safely with awareness of the full project
- Managing git: staging, committing, creating PRs with context
- Connecting to external tools via MCP (databases, browsers, APIs, Slack)
- Running in CI/CD for automated code review and issue triage
- Processing massive documents that break other tools
What sets it apart:
- Full filesystem access. It reads and writes real files in your real project.
CLAUDE.mdfiles act as persistent project instructions. Set your conventions once, and every session follows them.- Hooks let you enforce rules automatically (lint on save, format on commit).
- Multi-agent support. Spawn sub-agents for parallel work on different parts of a task.
- Works across terminal, IDE, desktop app, web browser, and mobile.
- No context window ceiling in practice. It manages its own context, compresses when needed, and keeps working.
Limitations:
- Usage limits depend on your plan. On Max plans ($100-200/month flat), you get generous usage without per-token billing. On API plans, extended sessions on large codebases can run $5-20+ each, and a heavy day can hit $50-100. Know which plan you're on and monitor accordingly.
- It follows conventions better when you tell it what they are. Without a
CLAUDE.mdfile, it'll guess at your patterns and sometimes guess wrong. Ten minutes of setup saves hours of corrections. - It takes action. This is both its strength and its risk. It will edit files, run commands, and commit code. Review diffs before accepting changes, especially on production code. The approval prompts exist for a reason.
- Vague instructions produce vague results. "Fix the bug" gets you a guess. "The login form throws a 403 after the session expires, check the token refresh logic in auth/middleware.ts" gets you a fix.
- The terminal is the barrier. If you've never used a command line, the first 30 minutes will feel unfamiliar. Push through it. Claude Code talks to you in plain English, and the terminal is just where it lives.
Getting started (5 minutes):
- Install:
curl -fsSL https://claude.ai/install.sh | bash - Navigate to any project:
cd your-project - Start:
claude - Try: "Explain the architecture of this codebase"
- Then: "Write tests for the auth module, run them, and fix any failures"
We built RIVER's entire AI platform with Claude Code. Not as a prototype. In production. Recently we needed to summarise an 800-page government report into a 20-page executive brief. Claude and Cowork couldn't handle it. Claude Code processed the entire document, structured the summary, and delivered it in one session. It's the tool with the fewest limits.
Isaac Rolfe
Managing Director
Why We Think Everyone Should Learn Claude Code
This might sound counterintuitive. Claude Code is a developer tool. Why would a COO or a strategy lead care?
Because Claude Code is where the ceiling is highest. Claude chat has context limits. Cowork needs your desktop running. Claude Code can process an 800-page PDF, manage a 200-file project, run for hours, connect to any external system via MCP, and operate on a schedule in CI/CD without anyone watching.
The barrier is the terminal. It looks intimidating if you've never used one. But here's the thing: Claude Code talks to you in plain English. You type what you want. It does the rest. The terminal is just where it lives.
If you're a leader who wants to understand where AI capability is actually heading, spend an afternoon with Claude Code. You don't need to write code. Just point it at a folder of documents and ask it to do something useful.
How to Choose
The decision is simpler than it looks:
| Question | Answer | Use |
|---|---|---|
| Do you need to think something through? | Yes | Claude |
| Do you need a document, spreadsheet, or organised output? | Yes | Cowork |
| Do you need maximum capability with fewest limits? | Yes | Code |
| Do you need your team on the same AI? | Yes | Claude Team/Enterprise plan |
Start with Claude. Everyone should. It's free to try and you'll see the value in one conversation.
Move to Cowork when you notice yourself spending 30+ minutes on tasks you could describe in one sentence. File organisation, data compilation, report formatting. If you can describe the outcome, Cowork can do the work.
Invest in Code when you're ready for the most capable tool available. Whether you write code or not. The learning curve is one afternoon. The ceiling is the highest of any AI product we've used.
What to Watch Out For (with any AI tool)
These aren't Claude-specific. They apply to every AI tool on the market right now.
AI is confidently wrong. Every model hallucinates. The better the model, the more convincing the hallucination. Never trust a critical fact, a specific number, or a citation without checking it yourself. This is especially dangerous in legal, financial, and medical contexts where a wrong number has real consequences.
AI amplifies your direction. If you give it a bad brief, you get a polished bad output. AI doesn't push back on a flawed premise the way a good colleague would. It'll execute your instructions perfectly, even when the instructions are wrong. The quality of your input determines the quality of the output.
AI isn't private by default. Understand the data retention policies of any tool you use. Free tiers typically use your conversations for training. Paid and enterprise tiers usually don't, but read the terms. If you're working with client data, sensitive documents, or IP, check what happens to your input before you paste it in.
AI doesn't replace judgement. It's a tool, not a decision-maker. Use it to surface options, draft analysis, and accelerate execution. But the decision is still yours. The organisations getting the most from AI are the ones that pair it with experienced humans who know what good looks like.
Start small, then expand. The fastest path to "AI doesn't work for us" is trying to transform everything at once. Pick one workflow. Prove it works. Expand from there. The organisations we work with that succeed all start with a single, boring, high-volume process.
How RIVER Uses All Three
We're not writing about these products from the outside. We use them every day to build enterprise AI platforms for NZ organisations.
Claude runs our strategy work. Every client engagement starts with a Claude Project loaded with the brief, the discovery notes, and our delivery framework. When we're reasoning through architecture decisions, scoping a new module, or drafting a proposal, it's Claude.
Claude Cowork handles odd jobs where it's easier to delegate than to do manually. File organisation, quick data compilation, reformatting a document for a different audience. Tim uses it for pulling together client status updates from scattered files. It's useful, but honestly, we lean towards Claude Code for most things because the capability ceiling is higher and we're already comfortable in the terminal.
Claude Code is the backbone. Every line of RIVER's Catalyst framework, every client deployment, every animation on this website. It reads our conventions, follows our patterns, and ships code that passes our quality checks. When we needed to compress an 800-page government report into a structured 20-page executive brief, only Claude Code had the capacity. Claude and Cowork both hit their limits. Code processed the entire document in one session.
The three products aren't competitors. They're different tools for different work. But if we had to pick one, we'd pick Code. It's where the capability ceiling is highest, and it's getting higher every month.
Getting Your Team Started
If you're considering Claude for your organisation, start with a Team plan. It gives you shared Projects (your team's docs and context in one place), admin controls, and usage analytics. Most organisations start with 3-5 people on a Team plan, find their use cases over 2-3 weeks, and then decide whether to expand. For Claude Code, each developer needs their own subscription. We run our team on Max plans for predictable costs and higher usage limits.