• Claude Code

Seeing Claude Code in Data Analysis Action

•

  • 4 min read

Target Audience: Anyone wanting to see Claude Code in action

Tools: Claude Code, Python, Git

Claude Code Key Points:
• Can analyse data anomalies
• Writes code on its own to investigate
• Self-adjusts direction based on what it finds along the way

This article tells a story in pictures of working within Claude Code. The simplicity masks the power you have at hand.

I start off with a chart that showed data in a shape I did not expect. I put Claude Code into planning mode to find the root cause. In the end it, it was found and fixed.

Just imagine how many other things Claude Code could investigate, hypothesise, and test.

The Problem

All the code that I do is 100% written by Claude Code. I wanted to see my repositories over time and what they are comprised of. So, I created a tool to count and classify lines1 and visualised my youtube-to-xml repository2.

I write a lot of tests, so when I saw a lot of purple, I was pleased. But three peaks in September? No, I don’t remember being extraordinarily efficient. But where to start, the source, the transformations, the code?

Stacked area chart showing Git repository growth with three anomalous 40k+ line spikes in September 2025; demonstrates self-built analytics tooling and data anomaly detection.
Apparent repository growth over time with three suspicious peaks in September

Claude Code Sherlock Holmes — Finding the Root Cause

All of these screenshots are taken from within my codebase that creates these images. Think of each as a page in a storybook of what it’s like to work with Claude Code. And here is the first page where I am instructing Claude Code to make a plan:

Claude Code terminal in plan mode with structured prompt defining investigation scope and data constraints; demonstrates strategic AI tool usage and systematic problem-solving approach.
1 of 8: Setting up the investigation in Claude Code’s Plan mode

On the last line of the image 1, I ask Claude Code to ask me questions first. This is for two reasons: first, what is crystal clear in my head is seldom clear in my writing. Second, there are aspects I miss but Claude Code doesn’t. Q&A is a good cheat:

Claude Code generating four structured diagnostic questions with multiple-choice options about repository data and Git history; demonstrates collaborative AI problem-solving and iterative requirement gathering.
2 of 8: Q&A with Claude Code pre-investigation

With my answers, Claude Code takes a preliminary look around in the files and then crafted the investigation plan:

Detailed 7-step investigation plan covering CSV inspection, Git history analysis, and code logic verification with clear deliverables; demonstrates systematic debugging methodology and technical planning skills.
3 of 8: Claude Code’s proposed plan of attack — approved

It started executing the plan at speed, a premature celebration followed by a “Wait, this doesn’t make sense”:

Claude Code discovering test file with 6,362 lines then recognizing discrepancy with current 248-line state; demonstrates iterative debugging and self-correction in AI-assisted investigation.
4 of 8: Found a clue, but something doesn’t add up.

Claude Code adjusts and this is what makes it powerful. A minute later the first glimpse of the bug I missed. One which would have taken me far longer to logically realise:

Code bug identified in groupby logic summing commits per date instead of latest state, with correlation showing 15 commits on peak days; demonstrates root cause analysis and data aggregation debugging skills.
5 of 8: Root cause identified—commits erroneously summed per day, not latest state.

And now the full explanation, supporting data, and proposed fix:

Root cause analysis showing bug location, evidence table with 15x inflation on peak dates, and 3-step fix proposal; demonstrates complete debugging workflow from diagnosis to solution.
6 of 8: Complete root cause analysis with proposed fix waiting for me to confirm — YES!

“Ready to fix the bug?” — “accept edits on” — which means Claude Code just goes and nothing needed from me. YES!

Fixed!

And here it is, about 20 seconds later the fixed chart and a completely different story:

Side-by-side comparison of buggy chart with 40k-line spikes versus corrected chart showing steady 3k-line growth; demonstrates complete debugging cycle from problem identification to validated fix.
8 of 8: Before and after—same data, completely different story

A Sneaky Tweak:

Looking at the fixed “right” chart, I obviously didn’t do much August with a big straight flat. Well, I don’t want to tell that story. I’ll do a sneaky tweak and ask Claude Code to make a new chart on the same data:

Single-line instruction to Claude Code generating complete new chart module with commit-indexed visualization; demonstrates efficient AI delegation and rapid iteration once patterns are established.
Instructing Claude Code to make one more chart — index by commit, not date — 240 commits in total.

With clean patterns to follow it is surprising how simple your instructions can be. The chart was created in one go. I went on to get Claude Code to do a “peak analysis” between commits [140,160]. And there was too, the cost of good old backwards compatibility.

And that is the story of how Claude Code investigated a data problem, fixed it, and then made me a new chart so I could tell a better story.

The End.

Footnotes

  1. plot-py-repo — the tool I created with Claude Code to visualise Python repositories ↩︎
  2. youtube-to-xml — the repository visualised in this article, also created with Claude Code ↩︎