Claude is Anthropic's AI assistant, and it's genuinely different from other models in ways that matter for day-to-day use. It's less likely to make things up, it follows nuanced instructions more reliably, and it handles long documents better than most alternatives. If you've been using ChatGPT and want to try Claude, or you're brand new to AI assistants entirely, this guide covers everything you need to go from zero to productive.
Getting access
Go to claude.ai and create a free account with your email. The free tier gives you access to Claude Sonnet — a capable model that handles the vast majority of tasks well. If you need higher limits, longer context, or access to Opus (the most powerful Claude model), Claude Pro costs $20/month.
You can also access Claude through the API if you're building applications, but for this guide we'll focus on the chat interface.
Your first conversation
Type anything in the message box and hit Enter. That's it. Claude reads your message, processes it, and responds.
But what you type matters enormously. This prompt:
"Write something about climate change"
…will get you a generic 300-word summary. This one:
"Write a 150-word explanation of why permafrost thaw creates a feedback loop, written for a high school student who already understands basic carbon cycles"
…will get you something actually useful.
The three things that make the biggest difference immediately:
- Say what you want, not just what you're curious about — "Explain X" is weaker than "Explain X so I can Y"
- Give context — Who are you? What are you working on? Claude uses this
- Specify format — "In bullet points", "in under 200 words", "as a table" all work
The difference between Claude models
When you start a new conversation, you can pick which Claude model to use. Here's what each one is good for:
Claude Haiku — Fast and cheap. Good for simple tasks, quick lookups, short drafts. If you're building something and need speed, this is the one.
Claude Sonnet — The sweet spot. Great reasoning, long context, fast enough for regular use. Most people should use this by default.
Claude Opus — The most capable model. Use it for complex reasoning, difficult writing tasks, or anything where getting it exactly right matters more than speed.
For most everyday tasks — writing, analysis, coding help, answering questions — Sonnet is the right choice. Switch to Opus when Sonnet isn't getting you what you need.
Uploading files and images
Claude can read documents, analyse images, and work with data. Click the paperclip icon in the chat box to attach files.
What Claude can read:
- PDFs, Word docs, text files
- Code files in any language
- CSVs and spreadsheets (it can summarize, find patterns, answer questions about the data)
- Images — screenshots, diagrams, photos
Practical example: Upload a PDF of a research paper and ask "Summarise the methodology and list the three most significant findings in plain English." Claude reads the whole document and gives you a focused answer, not a generic summary.
One important note: Claude processes files in context, not persistently. If you start a new conversation, you'll need to re-upload files. There's no permanent file storage in the chat interface.
How to use system prompts
System prompts are instructions you give Claude before the conversation starts. They're like a briefing that shapes how Claude responds throughout the session.
In the Claude interface, click "Add a system prompt" at the top of a new conversation. You can write anything here — Claude will follow it for the entire conversation.
A simple but powerful system prompt:
You are a writing editor specialising in technical documentation.
Always:
- Use plain English, avoid jargon unless necessary
- Flag when something is unclear before rewriting it
- Suggest one alternative phrasing at the end of each edit
Keep responses concise.
Now every message in that conversation follows those rules automatically. You don't have to repeat yourself.
Good uses for system prompts:
- Set a persona or role ("You are a senior data analyst…")
- Define output format ("Always respond in JSON with fields: summary, action_items, risks")
- Set constraints ("Never recommend specific products. Always ask for clarification if the request is ambiguous.")
- Provide background context ("The user is a nurse. Use clinical terminology but explain any specialist terms.")
The system prompts guide on MasterPrompting.net covers advanced system prompt patterns in detail.
Getting Claude to do what you actually want
The most common frustration with Claude (and any AI model) is getting an output that's almost right but not quite. There are a few reliable techniques for closing that gap.
Be explicit about what "good" looks like
Instead of "Write me an email to my client," try:
"Write a follow-up email to a client who's gone quiet after we sent a proposal two weeks ago. Tone: professional but warm, not pushy. Length: under 100 words. End with a specific question that makes it easy for them to reply."
Ask for drafts, not finished products
"Give me three different opening paragraphs for this blog post. I'll tell you which direction I want to go."
This costs you less back-and-forth than iterating on a single draft you don't like.
Tell Claude what went wrong
"That's too formal. Rewrite it as if I'm talking to a colleague I know well, not a stranger."
Claude responds well to specific feedback. "Make it better" does very little. "Make it shorter and cut the hedging language" does a lot.
Use XML tags for structured tasks
If you're giving Claude a lot of context plus a task, use tags to separate them:
<context>
We're a B2B SaaS company selling project management software to SMBs.
Our average customer has 10-50 employees. Main competitors: Asana, Monday.
</context>
<task>
Write five LinkedIn post ideas that position us as an expert in async work,
not just software features.
</task>
Claude handles structured input like this very well. It doesn't get confused about what's background and what's the actual request. The XML tags lesson explains why this works.
Working with long documents
Claude's context window on Sonnet 4.6 goes up to 1 million tokens in beta — that's roughly 750,000 words. In practice, you can paste an entire book, a full codebase, or months of meeting notes, and Claude can reason across all of it.
Practical applications:
Contract review — Paste a legal document and ask "What are the three riskiest clauses for the vendor? Flag any non-standard terms."
Code review — Paste a full file or function and ask "Find potential bugs. Focus on edge cases and error handling."
Research synthesis — Paste multiple papers or articles and ask "What do these sources agree on? What do they contradict each other on?"
Meeting notes — Paste a week of Slack threads or meeting transcripts and ask "What decisions were made? What's still unresolved?"
One tip: for very long documents, tell Claude what to focus on. "Read this report. I only need to understand the financial projections — ignore the marketing section."
What Claude is genuinely good at
- Writing and editing — First drafts, rewrites, tone matching, grammar, structure
- Summarising — Long documents, research papers, transcripts
- Code — Writing, debugging, explaining, reviewing — Claude is excellent at code across most languages
- Analysis — Breaking down arguments, finding logical gaps, comparing options
- Brainstorming — Generating ideas, variations, alternatives
- Structured output — JSON, tables, formatted reports
What to be careful about
- Facts and figures — Claude can hallucinate specific statistics, dates, and citations. Always verify factual claims from an external source, especially for anything you're publishing or presenting. The hallucinations lesson explains why this happens and how to reduce it.
- Recent events — Claude's knowledge has a training cutoff. For current news or recent releases, use a search tool alongside Claude.
- Confident-sounding errors — Claude doesn't always flag its uncertainty. If something sounds too neat or convenient, double-check it.
Five prompts to try right now
-
Rewrite something you've written — Paste a paragraph and ask "Rewrite this to be 30% shorter without losing the key points."
-
Explain something complex — "Explain how transformer models work to someone who understands Python but has never studied ML."
-
Debug your code — Paste a function that's not working and ask "What's wrong with this? Explain the bug before suggesting a fix."
-
Prepare for a meeting — "I have a meeting with [person/company] about [topic]. What questions should I ask? What objections might they raise?"
-
Stress-test your idea — "Here's my plan: [paste plan]. Give me the strongest three arguments against it."
The best way to learn Claude is to use it. Start with tasks where you already know what a good answer looks like — that way you can calibrate quickly. Once you have a feel for what it does well, the beginner prompt engineering track will give you the vocabulary and techniques to push further.
