Claude Code: The Agent in Your Terminal
Everything in this course — the loop, the context, the tools — shows up in one place at once: an agent that lives in your terminal and works inside your actual codebase.
Claude Code is the loop, given hands. It doesn't talk about your files — it opens them, changes them, runs them, and reads what happened.
The last lessons drew the line between a chat, which answers, and an agent, which acts, observes, and decides in a loop. Claude Code is that agent made concrete. It runs in your terminal — the plain text window from the CLI course — and instead of describing what you should do to your code, it does it, then looks at the result and keeps going. Four tools carry most of the work: it reads files (so it reasons about your actual code, not a guess), edits files (the same changes you'd make by hand), runs commands (your tests, your build, your git), and reads the output — the error, the pass, the diff — which is the “observe” that drives its next move.
One request, traced through the loop
Say I type one line. A chat would tell me what might be wrong; Claude Code runs the loop, and you watch each turn:
Read that top to bottom and you're watching act–observe–decide turn the wheel. Nobody handed it the answer; it found the bug the same way you would — by looking at what the code actually did.
Working with it well
A few habits separate a good session from a frustrating one. Explore, plan, code, commit: for anything non-trivial, let it research and propose a plan before it writes, so it doesn't confidently solve the wrong problem — skip the ceremony only for one-line diffs. Keep a CLAUDE.md with the project's standing context, but keep it short: for every line ask “would removing this cause a mistake?” — a bloated file gets ignored. Manage context aggressively (the last lesson's desk applies here too): clear between unrelated tasks, and name and resume sessions the way you'd branch. And for real confidence, have a fresh agent review the diff — a second pass, with context it didn't write, told to report real gaps rather than style preferences.
Where it breaks
Two opposite ways. Treating it like autocomplete — three vague words, expecting it to read your mind (everything from lesson one still applies; it's an agent you direct). Or pointing it at an enormous, fuzzy task and walking away. By default it asks before consequential moves; keep that leash short until it's earned slack, and stay a pair of eyes in the loop.
Try it yourself
In a small project you don't mind poking, give Claude Code one concrete, checkable task — “add a test for this function and make it pass.” Don't look away. Watch it run, read, edit, and run again, and notice the moment it reads an error and changes course. That moment is everything this course has been building toward: context, the loop, and a collaborator you direct — now working with its own hands.
Grounded in Anthropic's Claude Code documentation and Building Effective Agents.