What is Claude Code, and what can it do with data?

Claude Code is an AI agent from Anthropic that works right inside a folder on your computer. You give it a request in plain language, like "download NASA's exoplanet catalog and tell me how many planets there are by discovery method", and it does the work on your files: it downloads the table, writes a short program to read it and count the rows, runs that program on your machine, checks what came out, and hands you both the result and the steps it took.

That's the difference from a chat window. When you copy a chunk of a table and paste it into a ChatGPT or Claude conversation, the chat only sees the rows that fit in your message; ask it for a total or a percentage and it may try to work it out on the fly or estimate it from memory. Claude Code opens the whole file from your folder, whether it's a CSV with tens of thousands of rows or an Excel workbook, does the math by actually running code, and saves new files to your disk: a clean table, a summary, or an Excel workbook with formulas you can open and check.

The three pieces: model, harness and context

To understand what happens when you ask Claude Code for something, it helps to separate three pieces that work together in every session:

• The model: Claude's language model. It reads what's in front of it, reasons about the next step and writes code or answers. On its own, a model only takes in text and gives back text: it has no hands to open folders or run programs on your computer.
• The harness: the program that wraps the model and gives it real tools. Claude Code is that harness. When the model decides it needs to see a file's columns or run a count, it asks the harness; the harness asks for your permission when needed, runs the tool on your computer and hands the result back to the model.
• The context: everything the model has in view at that moment inside its context: your request, the file names and file contents it has already read, the output of the commands it ran and any instructions you've left in the folder. Of your work, the model only knows what has made it into that context, and when you close the session, the conversation doesn't carry over to the next one on its own.

With those three pieces, the agent works in short loops until it has solved what you asked:

Picture that loop with the exoplanet table you'll use in this course. You type your question. In the first loop, the model asks to look around the folder and the harness reads the first lines of the file to see what the columns are called. In the next loop, the model writes a short script to count the rows, the harness runs it and returns the result: there are 40,350 rows, but several of them belong to the same planet. If a command fails, or the rows need grouping a different way, the model reads the message that came back and takes another loop. Only once it has checked the result does it answer you on screen.

If you're still unsure what the model does and what the harness does when you ask for a calculation, ask your Maestro for an example.

How it reaches your data

An agent reaches your data in three ways, with the permissions you give it:

• Files in your folder: Claude Code only sees the folder where you opened the session. In this course you'll create an exoplanetas folder to hold NASA's catalog, the Excel workbooks it builds and your notes.
• Commands on your computer: Through the harness, it runs commands to download a table from the internet, look at columns or join files. Before it creates or changes a file or runs a command, it asks for your permission.
• Connectors and credentials: To query databases, shared spreadsheets or APIs at work, the harness uses connectors such as MCP or your team's credentials, without pasting passwords into the chat. Here you'll only use exoplanetas.

You can open the same Claude Code in the Claude desktop app (the Code tab), in the terminal with claude, at claude.ai/code, or in VS Code and JetBrains. The course is free; to use Claude Code you need a paid Claude plan or Anthropic API credits, because the free plan doesn't include it.

What Claude Code does and what you bring

With tables, Claude Code does the mechanical work fast: reading a whole CSV or Excel file, cleaning columns, filtering, grouping, computing percentages, joining files or building an Excel workbook with live formulas, all while showing you which commands it ran.

What isn't in the cells is what you bring:

• What each row stands for: the table's grain. NASA's catalog has 40,350 rows for 6,451 planets because each row is a publication, and one planet shows up as many as 34 times; depending on what you count, the same question gives 74.2% or 89.8%.
• What's in and what's out: which rows are tests, cancellations, duplicates or incomplete records.
• Which number you sign off on: which count answers your report and how to check that it adds up.

In the next modules you'll run your first session, check what it ran, and save your rules in an instructions file (CLAUDE.md) and in a skill, so every session starts out knowing your table. If you want to know whether the tables at your job are a good fit for Claude Code, tell your Maestro which files you work with.

Try it

To decide whether a task from your week belongs in a chat or in Claude Code, ask yourself: does it need to read or create whole files on my computer? And do I need the result to come from code that ran over every row?

Answer the question below first. Then think of two tasks from your week: one for a chat and one for Claude Code in a folder of yours, naming the input file and the file or verified count you want back.