How Does Claude Code Work? A Look Under the Hood

Claude Code works as a loop: it gathers context about your project, takes action with tools, verifies the result, and repeats until the task is done. Inside there are two parts — a model that reasons and tools that act — wrapped in what the docs call an agentic harness. Today I'm popping the hood, because once you see the mechanism, you stop treating it like magic and start steering it like a machine.

No computer science ahead, I promise. If you can follow how a dishwasher cycles — fill, wash, rinse, check — you can follow this.

What actually happens when you press Enter?

Your request kicks off three phases that blend into each other: gather context, take action, verify results. Ask a question about your project and the loop might stop at phase one. Ask for a bug fix and it cycles through all three repeatedly — reading files, editing, running tests, reading the failures, editing again. The key insight: Claude decides each next step based on what the previous step taught it, chaining dozens of actions and course-correcting along the way. It's not executing a pre-written script; it's navigating.

What tools does Claude Code have?

Without tools, an AI can only talk. Tools are what make Claude Code an agent: every tool use returns information that feeds the next decision. The built-in set falls into five families — file operations (read, edit, create), search (find files and text across the project), execution (run commands, start servers, use git), web (search and fetch documentation), and code intelligence (spot type errors after edits, via plugins). On top of that it can spawn subagents — helper copies of itself with their own clean workspace — and ask you questions when it's unsure.

What can Claude Code access on my computer?

When you run claude in a folder, it sees: the files in that folder and its subfolders (elsewhere only with your permission), your terminal, your git state, your CLAUDE.md instruction file, and auto memory — notes it saved for itself in past sessions. Your conversations are stored locally on your machine, as plain-text files under ~/.claude/projects/ — that's what makes resuming and rewinding sessions possible. And before Claude edits any file, it snapshots the current version first.

Can it break something? The two safety nets

Net one: checkpoints. Every file edit is reversible — press Esc twice to rewind to an earlier state, or just ask Claude to undo. Net two: permissions. By default Claude asks before editing files or running commands. Press Shift+Tab to cycle modes: default (asks), auto-accept edits (file changes without asking), and plan mode, where Claude only explores and proposes a plan without touching your files at all. The honest caveat from the docs: actions with external side effects — sending things, deleting things in online services — can't be checkpointed, which is exactly why Claude asks before those.

How do you use it well, knowing all this?

Three habits fall straight out of the mechanism. First, be specific upfront — name files, constraints, examples — because precise context means fewer loop iterations spent guessing. Second, give Claude something to verify against: expected results, a screenshot of the design, 'the page should show 5 items'. The verify phase is only as good as the target you define. Third, treat it as a conversation: when the first attempt misses, don't start over — say what's wrong and let the loop continue from there.

claude

Here's the kicker: tonight, give Claude Code one task and don't touch the keyboard until it finishes. Just watch the loop — what it reads, what it tries, how it checks itself. Ten minutes of watching the mechanism will improve your prompts more than any list of tips, because you'll finally know who you're talking to.