An agent, not autocomplete
Most people meet AI coding tools in one of two shapes: autocomplete that suggests the next line while you type, or a chatbot in a browser tab that answers questions but never sees your actual project. Claude Code is a third, more powerful shape: an agent living in your terminal.
| Shape | What it does | Who does the work |
|---|---|---|
| Autocomplete | Suggests the next few characters or lines as you type | You — it just speeds up typing |
| Chatbot | Answers questions; you copy-paste code back and forth | You — it advises from outside |
| Agent (Claude Code) | Reads your files, writes and edits code, runs commands like npm run build and git commit, and works in loops until a task is done | It builds — you watch and approve |
That last row changes everything. Because the agent can read your project, its suggestions fit your code instead of generic examples. Because it can run commands, it can build the site, see the error, and fix it — the loop a human developer runs all day.
You are the reviewer, Claude is the builder
Here is the mental model that makes everything else in this handbook click. Your job shifts from typing code to four things:
- Describe tasks precisely — say what should be true when the work is done.
- Review what gets built — read the changes like a senior developer reading a junior's work.
- Verify it works — run the build, click the page, check the result.
- Commit checkpoints — save known-good states so you can always step back.
One sentence to remember: you are the reviewer, Claude is the builder. Every habit, command and prompt template in this handbook flows from that sentence.
What a session actually looks like
You start Claude Code inside a project folder and type what you want in plain English — no special syntax:
$ cd my-portfolio
$ claude
> Add a dark-mode toggle button to the site header
# Claude reads your header file and styles,
# proposes edits and waits for your approval,
# then runs the build to prove nothing broke.
# You review the diff in VS Code and commit.The whole workflow of this handbook is that loop, made safe and repeatable: describe → approve → review → verify → commit.
Three truths to internalize on day one
- Context is everything — and costs money. Claude only knows what's in the current session, plus the files it's told to read. Every message you send carries the whole session with it. Small, focused sessions mean a smarter Claude and a lower bill — which is why you'll learn to clear the session between tasks.
- Files are memory, chat is not. Claude Code has zero memory between
sessions — by design. Anything worth remembering goes in a file (a
CLAUDE.md, adocs/folder). Chat evaporates; files persist. The Project Memory chapters build a whole system on this truth. - Trust, but verify — always. Claude is excellent and still sometimes wrong.
git diffbefore you commit, build before you merge, preview before you deploy. Non-negotiable, forever.
Real-world example — this website. Siazly is a static site built with the exact workflow this handbook teaches: every converter, calculator and learning page (including this one) was described as a task, built by Claude Code, reviewed as a diff, verified with a build, and committed as a checkpoint. Nothing here is theory.
Up next: install Claude Code and run it for the first time — one npm command, one login, about twenty minutes total.