Getting Started with Claude: Three Doors and a Four-Step Habit
Guide · George Pólya, How to Solve It~7 min read
The first time you open Claude, you will probably make the same mistake almost everyone makes: you treat it like a wish-granting machine. You type "build me a website" or "fix my essay," get back something vague, and quietly decide the tool is overrated. The tool is fine. The ask wasn't. This guide gets you past that first wall. By the end you will have signed in through whichever of Claude's three doors fits your task, run a real first session, and picked up a four-step habit — borrowed from a 1945 math book — that turns a flat answer into one you can actually use.
Difficulty · BeginnerTime · ~18 minPrereq · a computer online + one real task
Before you start
Have one real task in hand — a question you keep meaning to research, an essay to tighten, a script that keeps crashing. The whole guide works better with something concrete than with "let me just try it."
A browser is enough for the first door. The coding door also wants a terminal and Node.js 18 or newer; the API door wants a little Python.
Bring the right framing: Claude is a collaborator you brief, not an oracle you query. That one shift helps more than any clever trick.
Work out what you actually want — then pick the door
Before you type anything, answer the question George Pólya put at the top of his problem-solving method: what is the unknown? Not "I want help with my code" but "I want this function to stop crashing when the input is empty." A blurry ask earns a blurry answer; a sharp one gives Claude something to aim at. Once the task is clear, the door more or less picks itself, because Claude has three and they suit different work. claude.ai is a chat in your browser — best for questions, writing, research, thinking out loud. Claude Code is a coding partner that lives in your terminal and can read and edit the files in a project. The API is for when you want Claude inside software you are building yourself. Match the door to the task, and the rest of this guide is just walking the path you chose.
Try this
Write your task as one sentence before you start: "I want Claude to ___." If you can't finish that sentence, you aren't ready to prompt yet — you are still figuring out the problem, which is Pólya's first step and the one people skip most.
Door one — claude.ai, nothing to install
Open claude.ai in your browser and sign in with email, Google, or your work account. You land in a chat box, and that is the whole interface — type your task, press enter, done. The one choice worth understanding early is the model. Claude comes in tiers: Opus is the most capable for hard, multi-step work; Sonnet balances speed and capability for everyday tasks; Haiku is the fastest for simple, quick jobs. The default is a sensible starting point, so leave it there until an answer feels shallow, then reach for Opus. A free tier lets you start; paid plans raise the limits when you outgrow them.
Door two — Claude Code, invite it into your terminal
Claude Code is Claude as a command-line coding partner: point it at a project folder and it reads your files, runs commands, and edits code while you watch and approve. Install it one of two ways, then launch it from inside any project:
# Option A — native installer (recommended, no Node needed)
curl -fsSL https://claude.ai/install.sh | bash
# Option B — via npm (needs Node.js 18+)
npm install -g @anthropic-ai/claude-code
# then, inside any project folder:
claude
The first time you run claude it opens your browser to sign in to your Anthropic account — there is no key to paste. After that you are in a session: ask it to explain a file, hunt a bug, or make a change, and it proposes edits you approve before they land. Start read-only — "walk me through what this project does" — before you let it touch anything.
Door three — the API, put Claude inside your own program
The API is for builders: your own code sends text to Claude and gets text back, so you can wire it into an app, a script, or a data pipeline. Three small steps get you a first reply:
# 1. install the SDK
pip install anthropic
# 2. make a key at platform.claude.com → API keys → Create Key
# (copy it once — you won't see it again), then:
export ANTHROPIC_API_KEY="sk-ant-..."
# 3. your first call
import anthropic
client = anthropic.Anthropic()
msg = client.messages.create(
model="claude-opus-4-8",
max_tokens=1024,
messages=[{"role": "user", "content": "In one sentence, what is an API?"}],
)
print(msg.content[0].text)
Watch out
Keep the key out of your code and in an environment variable, the way the example does. It can spend real money, so treat it like a password — never paste it into a file you might commit or share.
Whatever door you took, prompt like you're solving a problem
This step decides whether Claude feels useful or disappointing, and it is the same behind all three doors. Pólya's four steps for solving a math problem map almost exactly onto briefing an AI. Understand: say what you actually want and why — "rewrite this for a non-technical reader" beats "improve this." Plan: hand over the context Claude can't see — the audience, the constraints, the file, an example of what good looks like — and for anything big, ask for an outline before the full answer. Carry it out: let it do the work, then read what came back instead of pasting it straight on. Look back: if it's off, don't restart with the same vague prompt — tell it exactly what was wrong and ask again. The single biggest upgrade for a beginner is the Plan step: most weak answers come from a prompt that assumed Claude could see what only you can see.
Look back — don't take the answer on faith
The last step is the one beginners skip and experienced users never do. Claude can be confidently wrong: it will state a fake statistic, invent a citation, or describe a function that doesn't exist, all in the same fluent voice it uses when it's right. So read its output as a strong first draft, not a verdict. Check anything that matters — a number, a name, a line of code — against a source you trust, or ask Claude to show its reasoning and its sources so you can see the seams. Run the code before you ship it. None of this is distrust of the tool; it is how you work with a fast, occasionally overconfident collaborator — gratefully, and with your eyes open.
Common mistake
The name for a confident fabrication is a "hallucination," and no model is free of them. The fix isn't a cleverer prompt — it's the habit of verifying. For a fact, ask "what's your source?" and check it. For code, run it. For anything high-stakes, keep the final judgment yours.
One pass, end to end. Say you want to turn a messy spreadsheet of survey replies into a short summary. You open claude.ai, because this is a one-off thinking task, not a codebase. You write the task as a sentence — "summarize the top three complaints in this feedback" — and paste the actual data, plus what "top" means to you (most frequent? angriest?). Claude returns three themes; you read them and notice it merged two you'd keep apart, so you say so. The second answer is right. Then you spot-check two of the quotes against the file before the summary goes in front of anyone. Four steps, five minutes, and a result you'd stand behind.
Brief Claude like you're solving a problem: state the task in one sentence, hand over the context Claude can't see, let it work and read the result, then look back and verify. If it's off, loop back with specific feedback rather than restarting. Framework: George Pólya, How to Solve It. An independent beginner's guide, not official Anthropic material.
Check your work
Can you say in one sentence what you wanted Claude to do this time?
Did you pick the door that fits the task — chat, Claude Code, or the API?
Did your first session or first call actually run end to end?
Did your prompt hand over the context Claude couldn't see, or did you toss it a one-liner?
After the answer came back, did you look it over and verify the facts that matter?
When it missed, did you give specific feedback and ask again — or just give up?
The one line to keep
Treat Claude as a problem-solving partner, not a wish machine: state the problem, hand over the context, then look back and verify — only then is the answer one you can trust.
动手打字之前,先回答波利亚摆在解题法第一位的那个问题:未知量是什么?不是"我代码想要点帮助",而是"我要这个函数在输入为空时别再崩了"。含糊的提问换来含糊的答案;清晰的提问才给了 Claude 一个能瞄准的靶子。任务一旦清楚,门基本就自己选好了,因为 Claude 有三道,各管不同的活。claude.ai 是浏览器里的对话——最适合提问、写作、查资料、把想法说出来理一理。Claude Code 是住在你终端里的编程搭档,能读、能改一个项目里的文件。API 则是当你想把 Claude 嵌进自己正在做的软件里。把门和任务对上,这篇剩下的内容就只是走你选好的那条路。
实操提示
开始前,把任务写成一句话:"我要 Claude 替我把 ___ 做成。"如果这句话补不全,那你还没到该写提示词的时候——你还在弄清问题本身,而这恰是波利亚的第一步,也是最常被跳过的一步。
这一步决定了 Claude 到底好用还是让你失望,而且三道门背后是同一套。波利亚解数学题的四步,几乎能原样套到"给 AI 交代任务"上。弄清问题:说清楚你到底要什么、为什么要——"把这段改写得让外行也看懂"远胜过"把它改好点"。拟定计划:把 Claude 看不见的上下文交给它——读者是谁、有哪些限制、那个文件、一个"好答案长什么样"的范例——大任务还要先让它给个提纲,再出全文。执行:让它干,干完先读一遍,别原样贴出去。回顾:不对劲就别拿同一句含糊提示重来——告诉它具体哪儿错了,再问一遍。对新手而言,单项收益最大的就是"拟定计划"这一步:大多数差答案,都出在那句假设了 Claude 看得见只有你才看得见的东西的提示词上。
回顾:别把答案当圣旨
最后这一步,新手会跳过,老手从不跳。Claude 可能一本正经地错:它会说出一个假数据、编一条不存在的引用、描述一个根本没有的函数,而且用的是它说对时一模一样的流利口吻。所以把它的输出当成一份很强的初稿,而不是定论。凡是要紧的——一个数字、一个名字、一行代码——都去你信得过的来源那里对一对,或者让 Claude 把它的推理和出处亮出来,好让你看清接缝。代码上线前先跑一遍。这一切都不是不信任这工具,而是你该怎么用一个敏捷、快速、偶尔过于自信的协作者——心怀感激,同时睁着自己的眼。
はじめて Claude を開くと、たいていの人と同じ間違いをするだろう。願いをかなえる機械のように扱ってしまうのだ。「ウェブサイトを作って」「この文章を直して」と打ち込み、ぼんやりした答えが返ってきて、この道具は過大評価だと静かに結論づける。道具に問題はない。頼み方に問題があったのだ。このガイドは、その最初の壁を越えさせる。読み終えるころには、あなたの用事に合った扉——Claude には三つある——から入り、最初の対話を実際に動かし、そして1945年の数学書から借りた4ステップの習慣を身につけている。平凡な答えを、本当に使える答えに変える習慣だ。
何かを打ち込む前に、ポリアが問題解決法の冒頭に置いた問いに答える。未知のものは何か?「コードを手伝ってほしい」ではなく「入力が空のときに、この関数が落ちないようにしたい」だ。あいまいな頼みはあいまいな答えを招き、はっきりした頼みは Claude に狙う的を与える。用事がはっきりすれば、扉はほぼ自分で決まる。Claude には三つあり、それぞれ別の仕事を受け持つからだ。claude.ai はブラウザの中の対話——質問、執筆、調べもの、考えを声に出して整理するのに向く。Claude Code はターミナルに住むコーディングの相棒で、プロジェクトのファイルを読み、編集できる。API は、自分で作っているソフトウェアの中に Claude を組み込みたいときのものだ。扉と用事を合わせれば、このガイドの残りは選んだ道を歩くだけになる。
初めて claude を走らせると、ブラウザが開いて Anthropic アカウントへのサインインを求める——貼り付ける鍵はいらない。そのあとはセッションの中だ。あるファイルの説明、バグ探し、ある箇所の変更を頼むと、編集案をまず示し、あなたが承認してから反映する。まずは読むだけ——「このプロジェクトが何をするか案内して」——から始め、それから手を触れさせよう。
第三の扉·API:自分のプログラムの中に Claude を置く
API は作り手のためのものだ。自分のコードがテキストを Claude に送り、テキストを受け取る。だからアプリ、スクリプト、データの流れに組み込める。三つの小さな手順で、最初の返事が手に入る。
問題を解くように Claude に指示する:用事を一文で言い、Claude に見えない文脈を渡し、やらせて結果を読み、振り返って確かめる。外れていたら最初からやり直さず、具体的な指摘を添えて前のステップへ戻る。枠組み:ポリア『いかにして問題をとくか』。独立した入門ガイドであり、Anthropic 公式の資料ではない。
仕上がりチェック
今回 Claude に何をさせたかったか、一文で言えるか?
用事に合う扉を選んだか——対話、Claude Code、それとも API?
最初の対話か最初の呼び出しが、最後まで実際に動いたか?
プロンプトは、Claude に見えない文脈を渡したか、それとも一言放り投げただけか?
答えが返ったあと、見直して、大事な事実を確かめたか?
外したとき、具体的な指摘をしてもう一度頼んだか、それともあきらめたか?
覚えておく一文
Claude は願いをかなえる機械ではなく、問題を解く相棒として使う。問題を言い、文脈を渡し、そして振り返って確かめる——四つを終えて初めて、答えは信頼できる。