The difference from a conversational assistant fits in one sentence: instead of giving you code to copy, Claude Code opens your files, edits them and runs your commands. It is an AI agent in the strict sense, applied to working on a project.
Definition
Claude Code is an agent developed by Anthropic that works directly inside a folder on your machine. It reads files, writes new ones, runs commands and observes the result to decide what comes next.
The sequence is that of an Agent loop: you give a goal, it explores, acts, checks, corrects. You approve what matters.
The name is misleading: the tool is not limited to code. It works on any folder of files, which makes it useful for documentation, data or structured writing, not only for development.
What changes compared with a conversation
| Classic assistant | Claude Code |
|---|---|
| You describe the context | It reads your files |
| It proposes, you apply | It applies |
| It does not know if it works | It runs the tests and corrects |
| Predictable cost | Cost varies with the passes |
The last line deserves attention. Like any agent, consumption depends on the number of loop passes, hence on the difficulty of the task. A clear request on a narrow scope costs little, a vague request on a large project can cost a great deal.
Best practices
Narrow the scope. A goal covering a few files gives far better results than an instruction covering the whole project.
Review before approving. The agent genuinely modifies your files. Version control is not a precaution, it is a prerequisite.
Give it something to check against. Tests, a verification command, a way to know whether the result holds. An agent that can check its own work corrects itself.
Connect your tools through MCP. That is what lets it consult your data or services without you pasting everything into the conversation.
An agent running commands on your machine can delete or overwrite files. Work in a versioned repository, and be wary of Prompt injection arriving from outside content the agent might read along the way.
Frequently asked questions
Do you need to be a developer?
For code, yes, at least enough to judge the result. For tasks on text or data files, no: the logic is the same and reviewing is within your reach.
How much does it consume?
It depends on the number of passes and the size of the context read. A ceiling and monitoring of pass counts remain the best guardrail, as with any agent.
Do the files go to the provider?
What the agent reads is sent to the model for processing, as in any conversation. For sensitive files, check retention commitments before pointing it at them.
How do you learn to use it well?
By starting with tasks you can judge, to calibrate your trust before widening. Our Claude Code course follows exactly that progression, guardrails included.