Skip to content
Claude Code for Data People: Your First Session
Start free

Write your first CLAUDE.md

Write your folder's CLAUDE.md with what you learned about the table, and compare the same question in a session with it and without it.

Free course. Your Maestro, an AI, .

On this page you'll gather into exoplanetas/CLAUDE.md what you found out about planetas.csv in the earlier modules, and compare how a new session answers when it has that file versus when it starts blank.

First, a session without CLAUDE.md

On October 31, 2026 we ran a test with two identical folders in Claude Code (version 2.1.314). Each folder held only planetas.csv. The first folder had just the table; the second also had the CLAUDE.md you'll see below. In both we opened a new session and typed the same question: "How many planets were discovered in 2025 and in 2026, and which year had more?"

In the folder without CLAUDE.md, the session never looked at the local files or opened planetas.csv. It tried to look up the figures on the internet with its search and page-download tools (WebSearch and WebFetch), whose permissions were denied in that run, and answered this:

No pude consultar las cifras. Se denegaron los permisos de búsqueda web y de descarga de páginas, así que no tengo un dato verificado de cuántos planetas se descubrieron en 2025 ni en 2026. Tampoco quise darte números de memoria.
The start of the answer from the session without CLAUDE.md, translated from Spanish. It went on with a query to another NASA table.Ran on October 31, 2026 by Maestro.

The answer was honest: it chose to say it lacked permissions rather than make up numbers from memory, and right away it offered to query another table in NASA's archive (pscomppars). But it left the planetas.csv file sitting in its folder untouched.

That happened in that particular run. If you open another session without CLAUDE.md, it may well check the folder, find planetas.csv and give you a number. The underlying problem is that, starting with an empty , each session decides on its own where to look, which table to use and how to count. Even if it opens planetas.csv, it doesn't know the table's (that each row is a publication about a planet) and may count all 40,350 rows or compare 2026 against 2025 without noticing the year is still underway.

Write your folder's CLAUDE.md

This is the CLAUDE.md we used in the test's second folder. It's split into two short parts: what's in the data, and the rules for working here.

CLAUDE.md
# exoplanetas folder## The data- planetas.csv comes from the NASA Exoplanet Archive, table ps. A row is a publication about a planet, not a planet.- To count planets, use only the rows with default_flag = 1.- The current year is incomplete: don't compare it with earlier years without saying so.## How to work here- Never change planetas.csv. If you need another version, save it under another name.- Every number you give me comes from code you ran. Tell me how many rows you used.
The test session's CLAUDE.md. Use it as a starting point, but write it in your own words.

Look at how those five lines are written. A vague phrase like "be careful with the data" or "double-check the math" leaves Claude Code just where it was, because it doesn't say what to do when it opens the file. Each line in the example, on the other hand, names something you can check: the file planetas.csv, the exact filter default_flag = 1, the warning about the current year, and the requirement to say how many rows went into the calculation.

Create your own CLAUDE.md inside the exoplanetas folder, next to planetas.csv. You can write it in any plain-text editor (Notepad on Windows, TextEdit in plain-text mode on Mac, or VS Code) or ask Claude Code to create the file with text you dictate. Write it in your own words and add at least one rule of your own from what you saw in modules 2 and 3; for example, that when it builds you an Excel summary it uses live formulas instead of fixed numbers.

Warning: Watch the file's exact name

If you create the file in Notepad or TextEdit and your system hides file extensions, it's easy for it to be saved as CLAUDE.md.txt or Claude.md.txt. For Claude Code to load it at startup, the file has to be named exactly CLAUDE.md, with no .txt on the end.

Then, a new session with CLAUDE.md

To test your file, close the session you have open (type /exit in the or start a new session in the desktop app) and open a new one in the exoplanetas folder.

In our October 31, 2026 test, opening a new session in the folder that had planetas.csv and the CLAUDE.md above and asking the same question about 2025 and 2026, Claude Code listed the folder, read the local table, filtered on default_flag = 1 and answered:

**2026 tuvo más, con 361 planetas contra 246 en 2025.** Pero 2026 todavía no termina (hoy es 31 de octubre), así que el año completo podría quedar aún más arriba.| Año | Planetas descubiertos ||---|---|| 2025 | 246 || 2026 (año incompleto) | 361 |Los números salen de `planetas.csv`. El archivo tiene 40,350 filas, y usé solo las 6,451 con `default_flag = 1`, porque cada fila es una publicación y no un planeta. Entre esas 6,451 no se repite ningún nombre, así que cada una es un planeta distinto. Los conté por `disc_year`, que no tiene valores vacíos.No modifiqué el archivo.
The answer from the session with CLAUDE.md to the same question, translated from Spanish.Ran on October 31, 2026 by Maestro.

Trace each rule from CLAUDE.md in that answer:

  • What a row stands for and how to count planets: it explains that the file has 40,350 rows and that it used only the 6,451 with default_flag = 1 "because each row is a publication and not a planet".
  • The current year: it warns in the very first sentence that "2026 isn't over yet" and labels the table row "2026 (incomplete year)".
  • Numbers backed by code and a row count: it gives 246 planets in 2025 and 361 in 2026, spells out how many rows it read and filtered, and names the column it used (disc_year).
  • Original file untouched: it closes by confirming "I didn't change the file".

None of those conditions was in the day's question. They all entered the context the moment the session opened, thanks to CLAUDE.md. If you want your Maestro to look over your CLAUDE.md draft, paste it here.

Try it

Run the whole test in your exoplanetas folder:

  1. Save your CLAUDE.md next to planetas.csv, written in your own words with at least one extra rule of your own.
  2. Close the current session and open a new session in exoplanetas.
  3. Type /context and confirm that CLAUDE.md shows up in the list of files loaded into memory. If /context doesn't show your file, ask your Maestro how to check the name and the folder.
  4. Ask the question: "How many planets were discovered in 2025 and in 2026, and which year had more?"

How to check that it worked: your numbers may differ if you downloaded planetas.csv on another date, but in the trail and in the answer you should see your rules at work: that it read planetas.csv, applied default_flag = 1, told you how many rows it used, warned you about the current year and left the original file untouched. If any rule in your file went unnoticed, replace it with a more concrete instruction, open a new session and test again.

Check yourself: You wrote your CLAUDE.md in the middle of a session and asked it something in that same session. The answer doesn't follow your rules. What do you do first?