The Testing Academy · AI for QA Study guide
AI for QA · Study guide

Claude Code 101, and the ten things the quiz is really testing

Anthropic's Claude Code 101 is free, twelve short lessons, and ends in a quiz. This is the whole syllabus compressed to one page, in the course's own running order, followed by ten questions with answers covering every idea the quiz can draw on. Read it before the course to know what to watch for, or after it as the revision pass.

By Pramod Dutta, The Testing Academy. Written for the batches at The Testing Academy. The syllabus and lesson order come from the course page itself. Every technical claim, file path, command and flag was checked against the current Claude Code documentation at code.claude.com rather than written from memory, because this tool changes fast and older guides are already wrong in places. The ten questions are mine, written to cover the assessed material; they are not the course's own quiz questions.

01

The course, and how it ends

Claude Code 101 is Anthropic's beginner course for its coding agent. Free, self-paced, and structured as twelve short lessons in four sections, then a quiz.

Section Lessons
What is Claude Code? What is Claude Code? · How Claude Code works
Your first prompt Installing Claude Code · Your first prompt
Daily workflows Explore, plan, code, commit · Context management · Code review
Customizing Claude Code The CLAUDE.md file · Subagents · Skills · MCP · Hooks
Quiz Course quiz

Stated prerequisites: basic familiarity with a code editor and the command line, plus a Claude account on Pro, Max or Enterprise, or an API key.

On the certificate, be prepared for either answer. The course page does not advertise a certificate in its own copy; the only mention sits in the privacy FAQ, which says tracked progress "allows us to provide you with completion certificates". Anthropic has been rebuilding its academy platform, and reports differ on whether you now finish with a certificate or a badge. Finish the quiz and see what you get. The same ambiguity exists on the AI Fluency course, covered on that page.

The number of quiz questions and the pass mark are not published anywhere on the course page. The ten questions at the end of this guide are mine, written to cover every idea the syllabus puts on the table.

02

What Claude Code is, and how it reads your code

Anthropic's own sentence: an agentic coding tool that "reads your codebase, edits files, runs commands, and integrates with your development tools."

The part worth understanding is what it does not do.

Not this Index the whole repo, embed it, search vectors This Read files on demand, the way you would read decide act grep, glob, open a file what does this need? edit, run a command, commit A loop, not a single answer.
It uses the same tools you would, which is why it can work on a repo it has never seen.

It runs on five surfaces that share one engine: terminal, VS Code, JetBrains, desktop app, and the web at claude.ai/code. Because the engine is shared, your CLAUDE.md, settings and MCP servers work the same on all of them.

03

The loop, and why testers start ahead

An agent is a model running in a loop with tools. Claude Code is that loop made concrete:

1234 PromptGatherActVerify you state the goalread, grep, call toolsedit, run, installdid that work? not verified? back to gather, and try again You can interrupt with Escape at any point, add context, or steer.
Four parts to name: tools, the context window, permissions, and the fact that it can be wrong.

Step four is where a QA audience starts ahead of everyone else. The agent's reliability rests on whether it can tell that it succeeded. Give it a passing suite and verification is real; give it nothing and step four is the model asking itself whether it did a good job. The corollary is uncomfortable and worth sitting with: hand it a flaky suite and you have given it a broken verifier, and it will confidently ship nonsense. Whether the suite is trustworthy was always your job. It just became the load-bearing one.

04

Permission modes, and where to sit

Press Shift+Tab to cycle. The current mode shows under the input box.

Mode What happens Use it for
Default Asks before every file edit and every shell command Anything touching a real repo
Auto-accept Edits go through without asking; commands still ask A throwaway branch you can reset
Plan mode Read-only tools only. It researches and proposes, it cannot edit Anything you have not thought through yet

Plan mode is the one to actually learn. It takes your prompt, analyses the codebase with read-only tools, and hands back a plan you can revise before a single file changes.

The dangerous mode is not auto-accept. It is default mode once you have stopped reading the prompts and are approving on reflex. Either read them, or work on a branch you are willing to throw away.

05

Installing, and the first prompt

# macOS, Linux, WSL
curl -fsSL https://claude.ai/install.sh | bash

# Windows PowerShell
irm https://claude.ai/install.ps1 | iex

# or with a package manager
brew install --cask claude-code
winget install Anthropic.ClaudeCode

Then, in any project:

cd your-project
claude

You log in on first run. If ANTHROPIC_API_KEY is set, it skips the login prompt and asks you to approve the key instead.

Native installs update themselves in the background. Homebrew and WinGet installs do not, so those need brew upgrade claude-code or winget upgrade Anthropic.ClaudeCode occasionally. Easy quiz material, and a real cause of "why does my version not have that feature".

06

The daily workflow

The course names one workflow and builds the rest of the section on it.

01020304 ExplorePlanCodeCommit read before writingagree the approachnow let it writesmall, reviewed wrong approach? back to plan, not on through code Skipping straight to Code is the single most common way to waste a session.
Plan mode exists for step 2: Shift+Tab cycles into it, and Claude proposes before it edits.

Two habits from the docs that make the loop work: press Escape the moment it heads the wrong way, and give it a way to check itself, a test to run or expected output, so it catches its own mistakes before you do.

07

Context management

Every session starts with a fresh context window, and everything you do fills it.

Command What it does
/context Shows what is consuming the window right now, including which memory files loaded
/clear Starts fresh. Costs nothing, and is the right move between unrelated tasks
/compact Summarises the history to free space. Takes instructions: /compact focus on the test output
/usage Token usage and, on a plan, what is drawing against your limits
One session, filling up startup your conversation room left /clear Empty window, nothing carried over. Costs nothing to run. Use it when the next task is unrelated. /compact Summarises the history and keeps going. Reads everything it summarises, so it is a big request. Use it when you need continuity, not a fresh start.
Auto-compaction does this for you near the limit. Doing it deliberately, earlier, works better.

Likely quiz shape: what survives /compact? A project-root CLAUDE.md does, because Claude Code re-reads it from disk and re-injects it. What does not survive is anything you only ever said in conversation. That is the argument for writing a rule down rather than repeating it.

08

Review, then ship

Three things that remove real friction, and one correction.

Review with a subagent before you push. It runs in its own context window with fresh eyes and none of the bias of the agent that just wrote the code. Two rules: restrict it to read-only tools, because a reviewer that edits is a second author, not a reviewer; and check the subagent definition into the repo so the team reviews against one standard.

Coming back to a PR. When Claude opens a PR through gh pr create, the session is linked to it:

claude --from-pr 4127

Days later, when review comments land or CI goes red, that picks up with the context intact. Verified against claude --help on the current build.

A correction worth making, because it appears in circulating notes for this course. /commit-push-pr is not a built-in command. It does not appear anywhere in the Claude Code documentation. Claude will happily commit, push and open a PR if you ask it to, and if you want that as one repeatable step, wrapping it is exactly what the skills layer below is for. --from-pr is real; /commit-push-pr you would have to create.

09

The five ways to customise it

The last section of the course is five lessons that are really one question: where should this instruction live?

Layer Lives at Loads Reach for it when
CLAUDE.md ./CLAUDE.md, ~/.claude/CLAUDE.md Every session A fact that is always true: build commands, conventions, layout
Skills .claude/skills/<name>/SKILL.md, ~/.claude/skills/<name>/SKILL.md On demand A repeatable procedure. Invoke with /name, or Claude picks it up
Subagents .claude/agents/*.md, ~/.claude/agents/*.md When delegated to A job that would flood your context, or needs its own tool limits
MCP Server config Tools listed, schemas on demand Claude needs to reach Jira, Drive, a database, your own tooling
Hooks .claude/settings.json, ~/.claude/settings.json At lifecycle events It must happen every time, whatever Claude decides
Context: shapes what Claude decides CLAUDE.mdSkillsSubagentsMCP always loadedloaded on demandown context windowreaches other tools All four are context. Claude reads them and tries to follow them. None of them is a guarantee. Configuration: runs whatever Claude decides Hooks shell commands A hook is a real process at a fixed lifecycle event. Exit 2 blocks the action. If it must happen every time, it is a hook, not a line in CLAUDE.md.
The distinction the whole section turns on, and the most likely thing to be tested.

A few specifics worth memorising:

  • /init generates a starting CLAUDE.md by reading your codebase. If one exists, it suggests improvements rather than overwriting.
  • CLAUDE.md files stack rather than override. Every one from the filesystem root down to your working directory is concatenated, so the closest one is read last.
  • Keep CLAUDE.md under about 200 lines. Longer files eat context and get followed less reliably.
  • A subagent has its own context window, which is the whole point: the log output stays in there and only a summary comes back.
  • A skill's body loads only when used, so a long reference document costs nothing until it is needed.
  • MCP is an open standard for connecting AI tools to external data sources.
10

Ten questions, with answers

These are mine, not the course's, and they cover every idea the syllabus puts on the table. Cover the answers.

1. Does Claude Code index or embed your repository before it can work on it? No. It reads files on demand using ordinary tools, the same way a developer would: search, open, read. That is why it can start on a repo it has never seen with no setup step.

2. You installed with Homebrew and a colleague's version has a feature yours does not. Why? Homebrew and WinGet installs do not auto-update; only native installs do. Run brew upgrade claude-code.

3. Name the four steps of the daily workflow, and the one people skip. Explore, plan, code, commit. People skip explore and plan, going straight to code, which is the most common way to waste a session. Plan mode, on Shift+Tab, exists for step two.

4. What is the difference between /clear and /compact? /clear starts an empty window and carries nothing over, and costs nothing. /compact summarises the conversation so far and continues with the summary, and because it has to read everything it summarises, it is itself an expensive request. Clear between unrelated tasks; compact when you need continuity.

5. Where can a CLAUDE.md file live, and what happens when there are several? Managed policy location, ~/.claude/CLAUDE.md for you, ./CLAUDE.md or ./.claude/CLAUDE.md for the project, and ./CLAUDE.local.md for personal project notes. They are concatenated, not overridden, ordered from the filesystem root down to your working directory, so the closest file is read last.

6. An instruction in CLAUDE.md keeps getting ignored on a step that must never be skipped. What should you do? Make it a hook. CLAUDE.md is context: Claude reads it and tries to follow it, with no guarantee. A hook is a shell command at a fixed lifecycle event and runs regardless of what Claude decides. This is the single most testable idea in the course.

7. What is a subagent for, and what makes it different from just asking again? It runs in its own context window with its own tools and permissions, and returns only a summary. So the verbose output, the logs, the search results, stays out of your main conversation. Definitions live in .claude/agents/ or ~/.claude/agents/, and only name and description are required.

8. When is a skill the right answer rather than a line in CLAUDE.md? When it is a procedure rather than a fact, and when you do not need it every session. CLAUDE.md loads every time and costs context every time; a skill's body loads only when it is invoked with /name or when Claude judges it relevant.

9. What is MCP, in one sentence? An open standard for connecting AI tools to external data sources, so Claude Code can reach things like Jira, Google Drive, a database, or your own internal tooling.

10. Which of the five customisation layers is enforcement rather than suggestion? Only hooks. CLAUDE.md, skills, subagents and MCP all shape what Claude decides. A hook is code that runs at a fixed point, and a PreToolUse hook exiting 2 blocks the action outright.

11

The commands worth memorising

Command What it does
claude Start a session in the current directory
Shift + Tab Cycle default, auto-accept, plan mode
Escape Interrupt and steer
/init Generate a CLAUDE.md by reading the codebase
/context See what is filling the context window
/clear Fresh window, nothing carried over
/compact Summarise and continue; takes instructions
/usage Token usage against your plan limits
/memory Browse and edit your CLAUDE.md and memory files
/mcp See and manage connected MCP servers
/hooks Browse configured hooks
claude mcp add Add an MCP server
claude --from-pr 4127 Resume the session linked to that PR
12

Three things people get wrong

  • Treating CLAUDE.md as configuration. It is context, delivered as a message. If it must be enforced, it is a hook or a permission rule.
  • Never running /context. It tells you exactly what loaded and what is eating the window, and it is the first thing to run when an instruction seems to have been ignored.
  • Letting one session run all day. Everything sends the whole conversation each turn. /clear between unrelated tasks costs nothing and is the biggest single saving available.
13

Where to go next

The course sits at the start of a track. After it: Claude Code in Action, then Agent Skills, then Subagents.