What is CLAUDE.md, and what goes in it for a data project?
Understand why every session starts out knowing nothing about your data, and separate what goes in CLAUDE.md from what goes in a single request.
Free course. Your Maestro, an AI, .
In the earlier modules you found several rules about your exoplanetas folder: that planetas.csv has 40,350 rows because each row records a publication, that counting the 6,451 unique planets means filtering on default_flag = 1, that the original file stays untouched, and that the counts in resumen.xlsx should be live formulas. All of that was clear in those conversations. But when you close Claude Code and open a new session today, the program starts from scratch.
What you talked about yesterday, what it discovered exploring planetas.csv and the conditions you set for building the spreadsheet all stay in the old session. In a new session, the only thing the model has in front of it is its Glossary word: the instructions it loads on opening, what you type at that moment, the files it reads during the conversation and what its tools return. On top of starting clean every session, that space has a size limit.
Picture a new hire on your team who walks into the office every morning remembering nothing from the day before. They're fluent at analyzing tables and writing code, but every morning you'd have to explain again what a row in the sales report stands for, which column filters the current records and which master file must never be overwritten. The practical move would be to leave a short note on their desk that they always read when they arrive, before touching any file.
In Claude Code, that note on the desk is called CLAUDE.md. It's a text file in your folder that the program's Glossary word loads automatically at the start of every session. In general, this kind of document is known as an Glossary word; CLAUDE.md is the name Claude Code uses.
Where it goes and how to check it loaded
The file is saved in your project folder with that exact name, the first part in capitals and the extension in lowercase: exoplanetas/CLAUDE.md. It can also live in the settings subfolder, at exoplanetas/.claude/CLAUDE.md. When you open Claude Code in exoplanetas, the program reads that file before you type your first word.
To check which memory files loaded in your current session, type the /context command and look at the list that appears on screen. If you run /context in your session and aren't sure what the list shows, paste the output to your Maestro.
If you're working in a folder that doesn't have a CLAUDE.md yet but does have an AGENTS.md file (the equivalent name other agents like Codex read), Claude Code can read that one instead.
What goes in CLAUDE.md and what goes in today's request
To decide where each instruction goes, ask yourself one question: is this true in any session I open in this folder?
In a data project, the answer separates your table's standing rules from each day's one-off questions. A table's Glossary word, the required filters for counting without duplicates, warnings about incomplete periods and rules about how to deliver results apply today, tomorrow and next month. The specific cut you need for a meeting, on the other hand, changes with every task.
| Goes in CLAUDE.md | Goes in today's request |
|---|---|
What a row of each table stands for and how to filter what matters (like default_flag = 1) | Which specific question you want to answer in this session |
| Which original files must stay untouched | Which chart, table or .xlsx file you need to produce now |
| Known quirks of your data (the current year is incomplete, one branch reports late) | Who the result is for or what format today's deliverable needs |
| How you work: every figure comes from code that ran, says how many rows it used and leaves formulas in Excel | Which years, methods or categories you want to compare this time |
What's in the left column you know from knowing your work, and a file's headers rarely explain it on their own. Writing it down makes any new session count correctly on its first try.
Split a mixed-up request before you send it
Worked example · 3 steps
Imagine that at work you open a new session in a sales folder and you're about to type this whole message in one go:
"Use ventas.csv, but never change that original file. Remember that each row is one product line within a receipt and that returns are flagged in their column so you can leave them out when adding up sales. Today I need to compare this month's sales by branch."
Before you send it, split it: what moves to CLAUDE.md and what stays in today's request?
Work out the next step yourself, then open it to compare.
Warning: Keep your file short and keep checking
Claude treats CLAUDE.md as context it reads at startup rather than as settings the system enforces. If you stuff it with instructions from past tasks or long passages, the file fills up with noise and the model follows your rules less well; the Claude Code docs recommend keeping it under 200 lines. And since it's still context the model interprets, you keep the habit of checking formulas and totals before sharing any figure.
Why in data projects it pays to write it in your own words
Claude Code includes the /init command, which explores the files in your folder and drafts a first CLAUDE.md. In programming projects it's very useful, because the code's own config files say how the program is installed and run.
In a data folder something different happens: if you run /init inside exoplanetas, Claude Code will see planetas.csv and its column names, but the headers alone don't tell the story of how that table was collected or which official filter your team uses in reports. That business knowledge is yours, so on the next page you'll write your own CLAUDE.md and test how it changes a new session's answer. If you want to pin down which rules would go in the CLAUDE.md for the tables at your job, tell your Maestro which files you use.
Try it
Before you move on to writing the file on the next page, do this quick exercise:
- Open a Claude Code session in your
exoplanetasfolder and type/contextto see what your session loads at startup today. - Write down three rules you learned about
planetas.csvandresumen.xlsxin modules 2 and 3 that should hold in any future session. - Answer the next question to check the difference between standing context and the day's request: