How to use Claude: a task-first guide for beginners

Getting started, and what it costs

Claude is Anthropic's AI assistant. It runs in a browser at claude.ai and in mobile and desktop apps, and using it is the same in all of them: you type a request, it answers, and the conversation stays there so you can push back on the answer. Creating an account is free and takes a minute. Anthropic gates accounts by country rather than by payment, so if signup fails outright, check the supported regions list before assuming something is broken.

Start on the free plan even if you already intend to pay. Free is a working product rather than a demo. It drafts and edits, reads documents you upload, searches the web, remembers things you told it in earlier conversations, and lets you create up to five Projects. A week on free tells you two things a review cannot: whether you like the writing, and whether you personally run into the limits.

What you get	Free	Pro (about $20 a month)
Writing, editing, and everyday questions	Yes	Yes
Web search	Yes	Yes
Memory across conversations	Yes	Yes
Uploading documents and images	Yes	Yes
Projects	Yes, capped at five	Uncapped, with deeper retrieval over uploaded material
Choice of Claude models	A default model	More models, including the stronger ones
Claude Code, the terminal coding agent	No	Yes, at limited usage
Usage before you are asked to wait	A session window that resets every five hours	Substantially more, plus a weekly ceiling aimed at very heavy use
Priority access when demand is high	No	Yes

Prices are Anthropic's United States list prices as of August 2026 and differ by country, with heavier Max tiers above Pro. If you are weighing the upgrade specifically, is Claude free good enough has a one-week test and is Claude Pro worth it covers what the paid plan changes and what it still will not do.

What Claude is strongest at

Three things, consistently, and knowing them is most of what separates a good first week from a frustrating one.

Writing that needs less editing. Give Claude a rough memo, a difficult email, or a set of notes and the draft that comes back usually needs fewer fixes than the alternatives before you would send it. It holds a register, follows nuanced style instructions, and avoids the padding rhythms that make AI text easy to spot. This is the single clearest reason people choose it.

Reading long and dense material. Contracts, filings, research papers, board packs, interview transcripts. Claude is careful with qualifications rather than flattening everything into a tidy summary, and it is more willing than most assistants to say a document does not answer your question instead of inventing something that sounds like an answer.

Reviewing code rather than only writing it. Ask what could break, where the hidden coupling is, and which tests would actually prove the fix works, and the answers tend to be specific. It also pushes back on an approach instead of cheerfully implementing whatever you asked for, which is the useful behaviour in a reviewer.

Two more habits worth knowing early. Claude takes correction well: tell it the third paragraph is wrong and it usually adjusts properly rather than apologising and repeating itself with new adjectives. And it follows long briefs, so a nine-point list of constraints tends to come back with all nine honoured.

What it does not do: generate images, and voice is well behind ChatGPT and Gemini. Very large context work, whole codebases or a year of transcripts, is Gemini territory. Claude vs ChatGPT has the head-to-head version of all of this with prompts you can run yourself.

The prompt pattern that works

Anthropic's own prompt engineering guidance and the experience of most heavy users land in the same place: be explicit, give context, and say what the output should look like. Four parts, in this order.

Task. The verb, stated plainly. Draft, rewrite, summarize, extract, compare, review, critique, plan, explain. "Help me with this" is not a task. "Rewrite this paragraph so it is 40 percent shorter without losing the caveat about delivery dates" is.

Context. Who it is for, what they already know, what you are trying to achieve, and the raw material. Claude guesses less when you tell it more, and it is unusually good at using background that other assistants ignore. Include the awkward parts: the deadline that slipped, the person who will be annoyed, the constraint you cannot change.

Constraints. What it must avoid as much as what it must include. Length, tone, banned phrases, facts that cannot change, claims you cannot make. "Keep every number exactly as I have it" and "do not add flattery" each fix a specific failure you would otherwise edit out by hand.

Format. A table, a memo, an email with a subject line, a checklist, five bullets, a two-paragraph answer. Naming the shape turns a vague reply into something you can use, and it is the cheapest single improvement available.

Copy-paste skeleton: Your task is to [task]. Context: [audience, goal, background, source material]. Constraints: [tone, length, must include, must avoid, facts to preserve]. Format: [table/email/memo/checklist]. Before you write, list anything you need from me that is missing.

That last sentence earns its place. Asking Claude what it is missing before it drafts surfaces the assumption it was about to make, and it is faster to answer one question than to rewrite three paragraphs built on a guess. Prompt engineering for beginners goes deeper on the technique if you want it.

18 prompts to start with

Replace the bracketed parts, paste your own material where it says paste, and keep the ones that work. Saving the prompts that earn their keep is what turns Claude from a novelty into a tool you actually reach for.

Writing and email

1. Here is a rough version of what I want to say. Rewrite it so it sounds professional but still like a person wrote it. Keep every fact exactly as I have it. No flattery, no filler openers. Under 150 words. Draft: [paste].
2. Write a message telling [person] that [difficult thing]. Direct but not cold. Do not over-apologise, and do not add hedging that invites an argument about the fact itself. End with one clear next step and a date.
3. Rewrite this in the voice of the sample below. Match sentence length, directness, and how much it explains. Do not change any fact or the order of the argument. Sample: [paste]. Draft: [paste].
4. Tell me the three weakest things about this piece of writing. Quote each one. Do not rewrite it, and do not tell me what is good about it. Draft: [paste].

Documents and reading

1. Read the attached document and give me: the five things that matter most, anything that would cost me money or create an obligation, and anything important that is vague or missing. Quote the exact text for each point so I can check it.
2. Compare these two documents. Produce a table of every place they disagree, with the wording from each side and which one is more favourable to me. Do not summarize the parts where they agree.
3. From this transcript, extract every decision, every owner, and every deadline into a table. Mark anything that was discussed but not decided as open. Do not infer owners that were not named. Transcript: [paste].
4. Here is a report. Before summarizing anything, give me a map: what each section covers and how long it is. I will tell you which sections to read closely.

Thinking and deciding

1. I am deciding between [options]. Score them against these criteria: cost, speed, risk, reversibility, and effect on [group]. End with a recommendation and the strongest argument against it.
2. Argue the opposite of the position in this document as convincingly as you can, using only evidence that is actually in it. Then tell me which of your counterarguments I should take most seriously. Document: [paste].
3. Here is my plan. List the assumptions it depends on, ranked by how much damage it does if that assumption is wrong. Do not suggest improvements yet. Plan: [paste].

Code

1. Review this diff for correctness, edge cases, and regression risk. Prioritise things that could actually break in production over style preferences. Tell me which tests would prove the fix works. Diff: [paste].
2. Explain what this code does, where it is likely to fail, and what a new developer would misunderstand about it. Keep it readable for someone who has never seen this codebase. Code: [paste].
3. Here is an error, the relevant code, and what I expected to happen. Give me the three most likely causes in order, the smallest diagnostic step for each, and the safest fix. Do not refactor anything I did not ask about. [paste].
4. I am considering [approach] for [problem]. Argue against it. What breaks at ten times the current scale, and what would I regret in six months?

Learning and everyday

1. Explain [concept] to someone smart who has never encountered it. One analogy, then a concrete example, then the three mistakes beginners make. No jargon unless you define it in the same sentence.
2. I have [time] a day for [number] weeks to learn [skill]. Build a plan with daily practice, a checkpoint at the halfway mark, and one small finished project at the end.
3. Here is my task list and my calendar for tomorrow. Group the tasks into deep work, admin, quick wins, and things to drop, and tell me what will not fit. Do not be optimistic about how long things take. [paste].

Documents, Projects, and Artifacts

Three features change how Claude works day to day, and beginners usually find them late.

Upload documents instead of pasting them. Attaching a PDF, spreadsheet, or image gives Claude the actual file rather than whatever survived a copy and paste, which preserves structure and stops tables turning into soup. For long files, ask for a map of the document first, then ask for extraction from the sections you care about, then check the quotes against the original. That sequence catches more than asking for a summary in one go.

Projects hold context so you stop repeating yourself. A Project is a workspace with its own uploaded material and its own instructions. Put a style guide, a brand voice document, and three examples of good past work into one, and every conversation started inside it already knows them. Free accounts can create up to five; paid plans remove the cap and add deeper retrieval over the material you upload. For anything you produce repeatedly, this is the feature that changes your habits.

Artifacts render work beside the conversation. When Claude produces a document, a piece of code, or a small interactive thing, it appears in a panel next to the chat and updates in place as you ask for changes, rather than being reprinted as another wall of text. Iterating on a draft this way is much easier than scrolling back through a chat log to find the version you liked.

One habit that costs nothing and helps constantly: start a new conversation when the topic changes. Claude re-reads the whole conversation on every turn, so a thread that has drifted through four unrelated subjects is both slower and more expensive than four short ones, and the answers get worse as the irrelevant history piles up.

The limits, and how they actually behave

Usage is measured in tokens rather than in messages, which is the fact that explains almost every confusing experience people have. A document is re-processed on every turn of the conversation it sits in, so one question about a 200-page PDF discussed across thirty messages costs vastly more than thirty short separate questions. Two people who sent the same number of messages can hit completely different walls.

Free accounts get a session window that resets every five hours, and how far it stretches also depends on how busy Anthropic's servers are at that moment. Paid plans get substantially more, plus a weekly ceiling aimed at very heavy use. Anthropic does not publish a fixed message count for any tier, and it moves with demand, so treat any article quoting a specific number, including the numbers you will find in forum threads, as a guess.

What drains a session fastest, in order: long documents inside long conversations, extended thinking on hard problems (the reasoning tokens count even though you only see a summary), large code files pasted repeatedly instead of referenced once, and very long threads that never get restarted.

The fixes are behavioural and they work. Start a new conversation when the topic changes. Attach reference material to a Project once instead of pasting it into every chat. Send the twenty relevant pages rather than the whole report. When a session gets long, ask for a written handoff summary and start fresh with it. People who work this way rarely meet the cap; people who keep one enormous conversation open all day meet it constantly.

Two other limits worth knowing. Claude's built-in knowledge stops at its training date, so anything recent needs web search turned on or the source pasted in, and it will answer confidently either way. And a large context window is not a promise of perfect attention across all of it, which is why verifying quotes against the original matters. What is a context window explains the mechanism in plain terms.

When a different model is the better tool

A limit you hit while pushing Claude at work it is not built for is a wrong-tool problem, and more quota does not fix it. This table is the short version of where each job goes.

The job	Reach for	Why
A draft that has to sound like a person wrote it	Claude	Fewest edits between the first draft and something you would send
Reading a dense contract, filing, or research packet	Claude	Careful with caveats, and willing to say a document does not answer the question
Reviewing code and pressure-testing an approach	Claude	Specific about risk and coupling rather than agreeable
Twenty headline variants or five versions of a page	ChatGPT	Range and speed rather than polish
A whole repository, twenty files, or a year of transcripts	Gemini	Context window in a different class
An answer with links you can open right now	Perplexity or a search-first tool	Checkable in seconds rather than minutes
Generating or editing an image	ChatGPT, Gemini, or a specialist model	Claude reads images but does not generate them
Talking to an assistant hands free	ChatGPT or Gemini	Voice is not where Anthropic has invested

The pattern that beats picking a favourite is drafting with one model and reviewing with another. Write in Claude, paste the result into a different model and ask for the weakest claim, then take the criticism back to Claude to fix. That loop is covered properly in how to use multiple AI models together, and a broader task-by-task ranking is in the best AI to use.

Running that loop across providers usually means several accounts, several histories, and several bills at roughly $20 each. A multi-model workspace like Whizi puts Claude alongside GPT and Gemini behind one login and one conversation history, which suits people who genuinely switch. It is the wrong choice if you want Claude Code, Projects with deep retrieval, or first access to Anthropic's new features, all of which live in Anthropic's own product.