TL;DR
Claude Code is not a chatbot that types code at you. It works in a loop: it looks at your project, makes a plan, takes an action with a tool, checks the result, and repeats until the job is done. Once you can see that loop running, you know exactly when to step in, when to let it work, and why it does what it does.
It is a loop, not a single answer
The mental model that trips people up is thinking of Claude Code like a chatbot: you ask, it answers, done. What actually happens is a repeating cycle. Claude takes one step, sees what happened, and decides the next step based on the result. That cycle is what people mean by 'agentic', and seeing it clearly is the single biggest unlock for using the tool well.
The four moves in the loop
- 1Read - Claude gathers context. It reads the files you point it at, searches the codebase, and looks at errors. It is building a picture of the situation before acting.
- 2Plan - it decides what to do next. For a small change that is instant; for a bigger task you can make the plan explicit with plan mode so you approve the approach first.
- 3Act - it uses a tool: editing a file, running a command, searching. This is the key point - Claude does not just talk about the change, it makes it, using real tools on your real project.
- 4Check - it looks at the result. Did the edit apply? Did the test pass? Did the command error? The outcome feeds the next turn of the loop, which is how it self-corrects.
Then it loops. Check feeds back into read, and the cycle turns again until the task is complete. A ten-step task is just this same loop running ten times, each turn informed by the last.
Why the loop matters to you
Once you can see the loop, a lot of confusing behaviour makes sense - and you learn exactly where your leverage is.
- Your best moment to steer is between turns. When Claude finishes an action and is about to plan the next, that is when a quick correction changes the whole trajectory. Press Esc and redirect - do not wait for a wrong plan to finish.
- Good context in means good actions out. The 'read' step is only as good as what you give it. Point Claude at the right files with @ and it plans from facts instead of guesses.
- It can recover from its own mistakes. Because it checks results, a failed command or test is not a dead end - it is information the loop uses to try again. Paste errors back in and let the loop work.
- Plan mode is you inspecting the plan step before any action happens. That is why it is worth using on anything non-trivial: you approve the approach before a single file changes.
Common questions
Is Claude Code just ChatGPT that can edit files?
Not quite. The defining difference is the loop: Claude Code takes an action, checks the result, and decides the next step based on what actually happened in your project. It works with real tools on real files and self-corrects, rather than producing one block of text for you to copy.
When in the loop should I interrupt?
The highest-leverage moment is right after an action, when Claude is about to plan its next step. A correction there redirects everything that follows. Press Esc, say what you actually meant, and let it re-plan - correcting early is far cheaper than fixing a finished wrong approach.
Why does giving it the right files matter so much?
Because the 'read' step is the foundation of every plan that follows. If Claude reads the wrong files or has to guess, it plans from bad information. Pointing it at exactly the right files with the @ symbol means the whole loop runs on facts.
What is plan mode and where does it fit?
Plan mode makes the 'plan' step of the loop explicit: Claude thinks through the full approach and shows it to you before taking any action. You approve or adjust the plan first, so nothing changes in your project until you are happy with the direction.
Get the visual guide to the agentic loop
Get 650+ plug-and-play skills, MCPs & prompts, plus 5,000+ builders - $9/mo, cancel anytime.
Join the Club