vibecodingby qairuhub
0/ 1960 XP
ENҚАЗ
All labsLab 01
Level 120 min+80 XP

Claude Code: Your First Session

Install it, point it at a real folder, and make it prove it understood your code before it is allowed to touch a line of it.

After this module you can

install Claude Code, run a first session in Manual mode, make it explore a project and explain it back, and leave the session cleanly

You need
  • A terminal you can open and 20 uninterrupted minutes (Lab 00)
  • Git installed (Lab 00) — on Windows that is Git for Windows
  • A paid Claude plan, a Console API key, or a guest pass — read the first section if you have none

Most first sessions with an agent go badly for one reason: people ask it to build something. You are going to do the opposite. Install it, point it at a folder, and spend the whole first session making it read — then check whether it actually understood.

By the end you will have a working install, a session you can leave and resume, and one prompt worth keeping forever.

What this costs, and what to do if you cannot pay

Claude Code is not included in the free claude.ai plan. You need Pro, Max, Team, Enterprise or a Console API account (setup docs). The cheapest door is Pro at $20 a month, or $17 a month billed annually — the pricing page prints that as “$200 billed up front”. Pro sessions default to Sonnet 5; Max, listed as “From $100 Per month”, defaults to Opus 5 (model config).

An API key is pay-per-token: cheap for one evening, expensive as a habit. Anthropic’s own cost docs put real usage at about $13 per developer per active day, $150–250 a month. A subscription is not unlimited either — your allowance resets on a rolling five-hour window plus a weekly one, shared with Claude chat and Cowork.

Kazakhstan is on Anthropic’s supported countries list, for claude.ai and the API. If you ever see App unavailable in region, that is a country or VPN problem, not an account problem (troubleshooting).

$20Pro per month, or $17 billed annuallyclaude.com/pricing, 21 Sep 2026 ↗
$13typical API cost per developer per active dayClaude Code docs, costs ↗
$0Antigravity CLI Individual plan, weekly limitsantigravity.google/pricing ↗

One more route: ask your mentor about /passes, described in the commands docs as sharing “a free week of Claude Code with friends” from an eligible account. Which accounts qualify is not documented, so ask rather than assume.

  1. Install it with one command

    Pick the line for your shell. On Windows, a prompt showing PS C:\ is PowerShell; C:\ without PS is CMD. Pasting the wrong one is the single most common failure in this lab.

    macOS, Linux, WSL
    curl -fsSL https://claude.ai/install.sh | bash
    Windows PowerShell — no admin needed
    irm https://claude.ai/install.ps1 | iex
    Windows CMD
    curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

    That is the native installer; it updates itself in the background. You need neither Node nor WSL — native Windows works. Git for Windows is optional, but install it from git-scm.com anyway: with it, Claude runs its Bash tool through Git Bash instead of falling back to PowerShell (setup docs).

  2. Open a NEW terminal and check the install

    The old terminal does not know about the new binary. Open a fresh one.

    check
    claude --version
claude doctor

    claude --version prints something like 2.1.269 (Claude Code). claude doctor is a read-only diagnostic of your install and settings — nothing it does can break anything.

  3. Give it something real to read

    Never run claude in your home folder or on your Desktop — from there it can see and index everything you own. Always cd into a project first. Have a small project already? Use it. Otherwise build this one; it takes three minutes, and it is small enough that you can check the agent’s answer against the code yourself.

    PowerShell
    mkdir qairu-cards; cd qairu-cards; git init
    macOS, Linux, WSL
    mkdir qairu-cards && cd qairu-cards && git init

    Now create three files in your editor. words.js:

    const PAIRS = [
  { kk: 'кітап',  en: 'book' },
  { kk: 'терезе', en: 'window' },
  { kk: 'қала',   en: 'city' },
  { kk: 'су',     en: 'water' },
  { kk: 'дос',    en: 'friend' },
];

    index.html:

    <!doctype html>
<html lang="kk">
  <head>
    <meta charset="utf-8" />
    <title>QAIRU Cards</title>
  </head>
  <body>
    <div id="card">...</div>
    <script src="words.js"></script>
    <script>
      let i = 0;
      let flipped = false;
      const card = document.getElementById('card');
      const render = () => { card.textContent = flipped ? PAIRS[i].en : PAIRS[i].kk; };
      card.onclick = () => { flipped = !flipped; render(); };
      document.onkeydown = (e) => {
        if (e.key === 'ArrowRight') { i = (i + 1) % PAIRS.length; flipped = false; render(); }
      };
      render();
    </script>
  </body>
</html>

    README.md:

    # QAIRU Cards

A tiny Kazakh-English flashcard page. Open index.html in a browser.
Click a card to flip it. Use the Next button to move to the next word.

    Read that README once more, then open index.html in a browser and click around. Keep what you notice to yourself for now.

  4. Start the session in Manual mode

    inside qairu-cards
    claude --permission-mode manual

    On Pro, Max and Team plans, a normal claude now starts in auto mode, where a second classifier model reviews actions instead of asking you (permission modes). That is convenient and you will use it later. For your first four labs, force Manual: it reads only, and it asks before anything else, so you actually see what the agent wanted to do.

    On first launch a browser opens for login. Credentials are stored, so this happens once; /login switches accounts and /logout signs out. The first time you run Claude in any folder you also get a workspace-trust dialog.

    The header shows three things worth reading: the version, the current model, and the working directory. Check the directory. If it is not qairu-cards, press Ctrl+C twice and cd properly. The status bar underneath should read manual mode on.

  5. Learn the controls before you ask for anything

    Do these in order, inside the session, before you ask it anything real.

    • /help — every command. There are around ninety; you need about fifteen.
    • ? on an empty line — the keyboard shortcut panel.
    • ! git status — shell mode. You run the command yourself and the output drops into the conversation.
    • Shift+Tab — cycles permission modes. Press it a few times and read the status bar each time: manual mode on, accept edits on, plan mode on. Land back on Manual.

    Two keys to memorise rather than try: Esc interrupts Claude mid-action and keeps the work so far, and Ctrl+O opens the transcript viewer, which shows every tool call in detail. Everything in this list is from the interactive mode docs.

    Finally run /usage. It shows your plan bars and where the tokens went. If you are on a subscription, ignore the dollar figure — the docs are clear that it only means something for API users (costs).

  6. Make it explore before it edits

    Here is the prompt. Type or paste it exactly. Notice what it does: it forbids edits, it scopes the reading to one small project, it asks for a claim you can check, and it ends by making the agent ask you something.

    Explore before editClaude Code · Manual mode
    Do not change any files yet.

Read this project and answer in plain text:
1. What does it do, in two sentences?
2. List every file and say what each one is for.
3. Where does the data live, and where is it rendered?
4. Name one thing the README claims that the code does not actually do.
5. What would break first if I added 200 more word pairs?

Then stop and ask me one question about what I want to build next.

    You will see it search and read files. It gathers context on demand — you never paste code in yourself.

  7. Judge the answer with a rubric, not a feeling

    A good answer does four things. Check them off:

    1. Names all three files, and says words.js holds the data while index.html holds the markup and the behaviour in one file.
    2. Finds the README mismatch. The README promises a Next button. There is no button in the code — only the Right Arrow key. If it missed this, it skimmed.
    3. Answers question 5 with a mechanism, not a vibe: all pairs load at once, there is no shuffle, no progress is stored, nothing is paginated.
    4. Ends with a question, because you asked it to.

    The tell of a bad answer is fluent, confident prose that never quotes a file. So test it:

    Make it show its workClaude Code · Manual mode
    Which of your five answers are you least confident about, and exactly what did
you read to decide it? Give me the file name and the lines you relied on.

    The docs have a name for this failure: the trust-then-verify gap — plausible code that misses the edge cases. Their fix is to give the agent a check it can run, a test or a script or a screenshot, and their rule is one line: cannot verify it, do not ship it (best practices).

  8. /clear versus /compact — the difference that saves your session

    Run /context. It shows what is filling the window right now: the conversation, every file read, command output, memory files. Sonnet 5 and Opus 5 hold a million tokens and auto-compact at about 967K by default (model config), but quality still degrades as the window fills. A big window buys you time, not quality.

    The rule, one line each:

    • /compact summarises the conversation and frees space, keeping the thread. Steer it: /compact keep the file list and the README mismatch. Use it on the same task, gone long.
    • /clear starts a new conversation with empty context. It costs nothing. Use it when the next thing you do has nothing to do with the last thing you did.

    Do it in that order now. Compact with an instruction, run /context again and watch the number drop. Then /rename first-session so /resume can find this conversation tomorrow, and /clear. In the empty session, try /compact once more: it answers Not enough messages to compact. — proof that compacting is about a conversation, not about the tool.

    One rule of thumb from the docs is worth copying down: if you have corrected Claude twice on the same thing, do not correct it a third time. /clear and write a better prompt.

  9. Leave, and prove you can come back

    Type /exit, or press Ctrl+D twice. Then, from the same folder:

    back in the shell
    claude -c

    -c continues the most recent conversation in this directory. Your session is there, with the file list it built. Exit again. You are done.

Why a read-only first session is not a warm-up

It is the actual skill. Three of the five failure patterns the Claude Code docs name are failures of context, not of code: unrelated tasks piled into one conversation, correcting the same mistake until the window is full of failed attempts, and unscoped exploration that reads hundreds of files to answer a question you could have scoped in one sentence.

An agent that has read your project answers differently from one that is guessing, and the difference is invisible in the prose. That is the whole reason step 7 exists. Asking “what did you read to decide that” costs you one extra prompt and catches the most expensive class of mistake in this course.

One last thing beginners forget: the tool documents itself. Typing how do I do X in Claude Code? inside the session works, and it beats searching.

Quick check

You have spent forty minutes with the agent on the flashcard page. It now understands the files well. You want to start a completely different task — writing a deploy script. What do you do first?
Before you close the terminal0/7 done
Take it with you · prompt-packThe explore-before-edit prompt

The five-question reading prompt, the follow-up that makes the agent cite its sources, and the four-point rubric for deciding whether it actually understood your project.

You are done when

A fresh session has read your project and explained it back to you correctly without editing anything, and `claude -c` reopens that same conversation.