Claude Code Explained: What Anthropic's Free

Anthropic's tools have a reputation for being technically coherent but not exactly beginner-friendly. The company released a free "Claude Code in Action" course through Anthropic Academy—also available on Coursera with a certificate—and Ani, a Stockholm-based senior tech consultant with over a decade in enterprise fintech, walked through the entire thing and distilled it into a seven-minute tutorial. What she covers reveals something about how Anthropic wants developers to understand Claude Code, and what the security pitch actually rests on.

What a Coding Assistant Actually Does

The course starts with a problem that sounds basic but matters more than it appears: large language models can only take text in and return text out. That's it. So when you ask Claude to read a file, something else has to happen—the model can't open a file system by itself.

This is where the coding assistant layer comes in. As Ani explains it, "the code assistant will send an instruction to Claude and will ask Claude to ask information about this document." In practice, that means the assistant translates your natural-language request into a structured instruction—something like: if you want to read a file, respond with 'read file, name of the file.' The LLM then uses that scaffolding to interact with tools it couldn't otherwise access.

It's a clean explanation of the agentic loop that's become the architectural backbone of tools like Claude Code: perceive context, formulate a plan, take an action, report back. The course treats this as foundational rather than incidental, which is the right call. Anyone trying to use Claude Code without understanding this loop will eventually hit a wall they can't explain.

The Security Argument—And Its Limits

The most substantive claim in the course concerns security, and it's worth examining carefully because Anthropic has been making variations of it as a differentiator.

The argument goes like this: tools like GitHub Copilot Workspace perform code indexing, which means chunks of your codebase get stored on the vendor's servers. Claude Code doesn't do that. Instead, it runs general filesystem commands—like find—on your local machine and sends only the relevant file contents to the Claude API when needed.

Ani frames this plainly: "When you use Claude Code, Claude Code inside your computer will use general commands, for example, find to get the right information, and only the correct information will be sent to vendor where Claude API is hosted instead of just sending all the code base."

For enterprises handling sensitive financial or proprietary code, this distinction is not trivial. The difference between "we index everything upfront and store it" and "we query locally and send targeted content on demand" represents a meaningful architectural choice, and the security implications downstream are real.

That said, the argument deserves some scrutiny. The claim is about data minimization—Claude Code sends less of your code to Anthropic's servers at any given moment. It is not a claim that nothing leaves your machine. Every query still involves content going to Anthropic's API. Whether the threat model for a given developer or organization makes this distinction meaningful depends heavily on what they're building and for whom. This nuance doesn't undermine the course's point, but it's the kind of detail that matters before deploying any AI coding tool in a regulated environment—something worth checking alongside security gaps that have surfaced in Claude Code's development history.

Three CLAUDE.md Files, Three Different Scopes

The most practically useful section of the course covers what Anthropic calls CLAUDE.md files—a three-tier configuration system that shapes how Claude Code behaves across different contexts.

Project-level CLAUDE.md: Generated by running /init in a project directory. The command reads your codebase and produces a summary file capturing the project's structure, relevant files, and context. This is the file that gets shared across a team—it lives in the repo.

CLAUDE.local.md: Personal, not committed. This is where individual developers put preferences they don't want to push to the shared project config. Ani's example: if Claude is adding too many comments to your code, you can instruct it here to "use comments sparingly, only comment complex code." The instruction persists across sessions without polluting the team's shared context.

Global CLAUDE.md: Machine-wide. Instructions here follow Claude Code across every project on your computer.

The layering is sensible. It mirrors how most developers already think about configuration—project defaults, personal overrides, global preferences. Whether this three-file system is sufficiently discoverable for new users is a different question; the course at least makes the structure explicit in a way the raw documentation may not.

Running /memory inside Claude Code opens the relevant file for direct editing. Tagging specific files with the @ symbol followed by a file path lets you point Claude at exactly the right context without forcing it to scan the entire project—a performance and relevance optimization that Ani positions as habit-forming: "if you already know in which file is the information, then you can just tag the file."

What the Free Course Is—And Isn't

There's a category question worth sitting with here. This is Anthropic's official curriculum, taught through their own Academy platform and mirrored on Coursera. It is, by definition, a product tutorial. That doesn't make the information wrong—most of what the course covers is technically accurate and useful—but it does mean the framing consistently favors Claude Code over alternatives, and the security comparison with GitHub Workspace is made without acknowledging the conditions under which the comparison holds or doesn't.

Ani herself is clear about the context: she builds AI workflows for asset managers and pension funds, and she's filtering the course through that lens. Her enterprise fintech background makes the security framing more relevant than it might be for, say, an indie developer building a side project in a public repo. The course's claims land differently depending on who's asking.

The free access point is real. Anthropic Academy requires no payment for the core curriculum, and the Coursera version adds a certificate for those who need to demonstrate completed training. Given how rapidly Claude Code's feature set has been expanding—git worktrees, security scanning, desktop previews have all landed recently—a structured introduction to the fundamentals has genuine utility, even if the more advanced capabilities outpace what the beginner course covers.

The Broader Pattern

What this course reflects, more than anything, is the current state of agentic AI tool adoption: we're in a phase where the infrastructure is real and functional, the security questions are partially answered and partially open, and the onboarding materials are still catching up to the complexity of what developers are actually deploying.

Anthropic making this curriculum free and accessible is a distribution strategy as much as an educational one. Getting developers fluent in Claude Code's configuration patterns—especially CLAUDE.md files and the /init workflow—creates stickiness that's hard to replicate through API documentation alone. Whether the quiet market capture this enables translates into durable enterprise adoption, or whether competitors build equivalent configurability into their own tools and close the gap, is the more interesting question ahead.

The course teaches you how to use Claude Code. It doesn't yet answer whether the architectural choices baked into it will hold up as autonomous permissions and agentic workflows push into more sensitive production environments. That's where the real policy and security conversation is headed—and where a seven-minute tutorial, however well-made, runs out of road.

Samira Okonkwo-Barnes covers technology policy and regulation for Buzzrag.