A skill that checks your table before it counts
Write the revisa-tabla skill, run it with /revisa-tabla, and confirm Claude Code loads it on its own when you ask it to count something.
Free course. Your Maestro, an AI, .
In Module 2 you checked planetas.csv step by step before trusting a count: how many rows it has, what each row stands for, whether the same planet shows up more than once, and whether anything is missing. Instead of typing those questions every time a new file lands on your desk, you can save them in a Glossary word so Claude Code runs the same check, in the same order, every time.
Create the folder and the SKILL.md file
Inside your exoplanetas folder, create a .claude folder, inside it a skills folder, and inside that a revisa-tabla folder. At that Glossary word, save a file named exactly SKILL.md with the instructions below.
If you'd rather not make the folders by hand in your file browser, ask Claude Code in your current session to create .claude/skills/revisa-tabla/SKILL.md with this content. That way Windows or Mac can't sneak a hidden .txt onto the end of the name.
---name: revisa-tabladescription: Checks a data table (CSV or Excel) before counting or averaging anything with it. Use it when I ask you to analyze, summarize or count something from a new table.---Before you calculate any number with the table I point you to ($ARGUMENTS):1. Tell me how many rows and how many columns it has.2. Tell me what one row stands for. If the same thing (a person, a planet, a ticket) appears in several rows, say so and tell me which column tells them apart.3. Tell me which columns have empty cells, and how many.4. Show me five rows spread across the file: two from the start, one from the middle and two from the end.5. Don't change the original file.Finish with one sentence: "To count X in this table, count Y", where Y says which rows or which column to use.It's only a few lines, but this file has the full shape of a task spec for an agent, and it covers the basic profiling of any table:
Goal and trigger in description
The header between the
---lines tells Claude Code what the skill is called and when it makes sense to load it: before counting, summarizing or averaging data from a new table.Input and profiling questions
The
$ARGUMENTSvariable receives the file name. Steps 1 through 4 profile the table's quality: its size, what each row stands for, which column identifies unique records, how many empty cells there are, and five real rows spread across the whole file.An explicit constraint
Step 5 sets the task's safety line: inspect and read the data without touching the original file.
A finish line you can check
The last instruction requires closing with the sentence
"To count X in this table, count Y". One glance tells you whether the analysis is done and which rule it will apply.
Look closely at steps 2 and 4. Step 2 forces a check of the Glossary word before anyone adds or counts rows, and step 4 asks for a Glossary word with rows from the start, the middle (around line 20177 in planetas.csv) and the end (lines 40350 and 40351), so you're never judging the file by its first few lines alone.
Common mistake: Two slips that keep Claude Code from finding your skill
The file has to be called SKILL.md, in capitals, and live inside .claude/skills/revisa-tabla/. If you save it straight into .claude/skills/ or name it revisa-tabla.md, Claude Code ignores it. Also make sure the very first line of the file is exactly ---, with no blank lines above it, and that another --- closes the header after description.
Call it by name
To try it directly, open a new session in your exoplanetas folder and type the command followed by the table's name:
/revisa-tabla planetas.csvIn the test session on October 31, 2026, with Claude Code 2.1.314, the agent read planetas.csv with Python, looked at lines 2, 3, 20177, 40350 and 40351 of the file, confirmed it hadn't modified the original, and answered point by point:
**1. Tamaño:** 40,350 filas de datos (sin contar el encabezado) y 5 columnas: `pl_name`, `hostname`, `disc_year`, `discoverymethod`, `default_flag`.**2. Qué representa una fila:** una fila no es un planeta. Es una fila por cada publicación o medición de un planeta. Solo hay 6,451 planetas distintos (`pl_name`), así que un mismo planeta aparece en varias filas. TrES-2 b aparece 34 veces y Kepler-10 b 25. Solo 937 planetas aparecen una vez.La columna que distingue la versión "oficial" es `default_flag`. Vale 1 en exactamente una fila por planeta (6,451 filas) y 0 en las otras 33,899. Lo comprobé: ningún planeta tiene dos filas con 1 ni se queda sin ninguna.**3. Celdas vacías:** ninguna columna tiene celdas vacías (0 en las cinco). Todas las filas tienen 5 campos.[...]Para contar planetas en esta tabla, cuenta las filas con `default_flag` = 1 (o los valores distintos de `pl_name`). Cuenta filas sin filtrar solo si quieres contar mediciones publicadas, no planetas.Everything you worked out over several turns in Module 2 shows up here in a single reply, and it ends with the counting rule you asked for.
Check that it loads on its own when you ask for a count
The payoff of a good description is that you don't have to remember to type /revisa-tabla every time. In a second test that same October 31, 2026, we opened another fresh session (with CLAUDE.md in the folder) and asked only:
"How many planets were discovered in 2025 and in 2026, and which year had more?"
Without anyone naming the skill, Claude Code saw that the question meant counting data from a table, loaded revisa-tabla by itself, ran the check first, and then answered: 246 planets in 2025 and 361 in 2026.
One detail in that second run matters: when it printed the five sample rows, its command's output got cut off, and Claude Code flagged it in its reply: "My command only printed a truncated view, so the sample is incomplete." A skill makes the agent follow your checklist and report where it fell short, but reading that warning and deciding whether to ask for the missing rows is still your job.
Fit the steps to your own tables
The SKILL.md on this page is a starting point, but your work tables have quirks NASA's catalog doesn't: dates in mixed formats, canceled invoices still sitting in the database, amount columns with and without tax, or branches that send their numbers a day late. When you bring this skill into a work folder, change or add steps so it checks whatever has made you redo a count before. If you want to write a step for the tables you use every day, tell your Maestro what usually goes wrong in your files.
Try it
Run both tests on your computer:
- Save
.claude/skills/revisa-tabla/SKILL.mdin yourexoplanetasfolder, open a new session and run/revisa-tabla planetas.csv. - Open another new session and ask a counting question without naming the skill, for example:
"How many planets were discovered in 2025 and in 2026, and which year had more?".
How to tell it worked: In the first test, check that the reply includes the 40,350 rows and 5 columns, the 6,451 distinct planets, TrES-2 b showing up 34 times, 0 empty cells, the five sample rows, and the closing sentence with default_flag = 1 or pl_name. In the second test, check that Claude Code says up front that it loaded revisa-tabla before giving you the totals of 246 and 361. If Claude Code says it doesn't recognize the command, or answers straight away without loading the skill, tell your Maestro which path you used and what you saw.