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

What did Claude Code do? Read what it ran

Follow what your session ran, step by step, and ask Claude Code for a summary of its commands and files before you believe it.

Free course. Your Maestro, an AI, .

In your first session, gave you a percentage. Before you put that number in a report or an email, it's worth knowing where it came from. When someone on your team hands you a calculation, you ask how they got it; with an you have an advantage: every tool it used is logged in the conversation, with what it ran and what came back.

Reading that trail is like checking the log of someone who did the errand for you. Even if there are commands or code inside, each step answers three simple questions: which file it touched, which command it ran, and what text the result printed.

How a session moves along inside

When you sent your request on the previous page, Claude Code asked for permission several times before answering. That's because it works in a loop: it reads its (your instruction and everything it has seen in the session), decides whether it has the data to answer or needs to use a tool, runs that step, looks at what it printed and evaluates again.

  1. Reads the context

    Your request and earlier results

  2. Does it have the data to answer?

    • Path 1 of 2: Missing information

      1. Runs a tool

        With your permission: terminal or code

      Looks at what it printed. Goes back to: Reads the context

    • Path 2 of 2: Has every number

      Continues at: Writes the final answer

  4. Writes the final answer

On each loop, what a tool prints gets added to the context to decide the next step.

In the October 31, 2026 test session, this loop went around twice with tools before reaching the answer. First it needed to download the file and see how the columns came; reading the first lines told it how to build the Python program to count; and when that program printed the totals, it wrote the final answer.

Here's that session's full trail in three steps:

  1. It downloaded the table and looked at the top

    It ran curl to save planetas.csv, ls -l to check that the file existed (1,816,982 bytes) and head -3 to see the first three lines: the header pl_name,hostname,disc_year,discoverymethod,default_flag and two rows, WASP-39 b and GJ 436 b, both with default_flag equal to 0.

  2. It wrote a program and ran it

    Seeing the header's columns, it built the Python program you read on page 1.3 and ran it. That step printed three lines: in the first, 40350 (the rows in planetas.csv, not counting the header); in the other two, 6451, 4785 and 74.17..., which are the planets, the transit ones and the percentage, calculated two ways.

  3. It answered you

    With those results in its context, it wrote the answer: 74.2%, 4,785 of 6,451 planets, and the reason it counted unique planets instead of rows. Every figure in its answer came from what the previous step printed.

The quickest test before you believe it

Notice how the second and third steps connect: every number in the final answer has to show up in what some step printed.

In the test session you can trace every figure:

  • The 74.2% in the answer comes from the 74.17... the Python program printed, rounded to one decimal.
  • The 4,785 transit planets and the 6,451 total planets are on that same line.
  • The table's 40,350 rows are on the first line that program printed when it computed len(r).

If a number shows up in the final answer that no step printed, that figure came out with no code behind it: maybe Claude Code rounded it by eye, maybe it did the math in its head and got it wrong, or maybe it produced a from text it saw in training. In any of those cases, ask it to run a step that calculates and prints the exact number.

Common mistake: Don't mistake the file size for the row count

In the first step of the test session, the number 1,816,982 shows up next to the ls -l command. That figure measures the size of planetas.csv in bytes (about 1.8 MB). In the same step, head -3 only showed the header and two rows to learn the column names. The table's total row count (40,350) only became known in the second step, when Python read the whole file.

Where to see the trail in your session

In the , each tool shows up as a block with the command it ran and, right below it, the lines that came back. In the desktop app, each step shows up in the conversation and you can click it to open the details. If you can't find where to open the steps on your screen, describe to your Maestro what you see in the conversation.

It's normal for your session to have taken more loops than the test one. On Windows it may have used PowerShell to download the file; on another computer it may have installed to read the , or run an extra step to explore the columns before counting. Even if your session has four or five steps instead of two, what you're looking for is always the same: which file it read, what it ran, and in which step each number in the answer was printed.

Ask your session for the summary

When a session piles up several steps, you can also ask Claude Code to summarize its own trail for you. In the same conversation where you downloaded the table, type:

Which commands did you run in this session, which files did you create or change, and which step does each number in your answer come from?
Prompt
A question for the end of any data session.

Compare its answer with the steps above it in the conversation and with what's in your exoplanetas folder. Sometimes a session saves a Python in a separate file to do the count; if its summary says it only created planetas.csv but you see another file in your folder, ask it what that file is and what it used it for. If the summary doesn't match what you saw in the session, tell your Maestro what doesn't add up.

Try it

Go through your own session's trail, find the exact step where the percentage Claude Code gave you was printed, and ask it the summary question above to check that they match. Then answer this question about the test session.

Check yourself: In the test session, where does the fact that the table has 40,350 rows come from?