Spent the weekend tinkering with Claude Science; hit many pitfalls but finally figured out how to use AI for delivery data analysis

Bottom line: Claude Science and similar data analysis tools are absolutely worth it. But beginners jumping straight in will likely get stuck at the first step 80% of the time. I spent an entire Saturday afternoon just to get a complete analysis flow working. If you also want AI to help you look at data, this zero-to-one tutorial should save you at least half a day.

Let's clarify one thing first: Claude Science isn't a tool where you ask "analyze this data" and it spits out perfect conclusions. It's more like an assistant helping you with exploratory data analysis. You give it data, and it helps find patterns, propose hypotheses, and create visualizations. But the prerequisite is that you must tell it what you want to see and how to ask.

I'll use Zeekr's July delivery data as an example. 35,837 units, up 111% year-over-year, hitting a new historical high. The number looks good, but as a researcher, what I want to know is: Is this growth structural or incidental? Which models are carrying the load? Are there hidden patterns behind the data?

Step 1: Prepare Data Files

Don't expect Claude Science to parse web pages directly. It needs structured data, like CSV or Excel. My first pitfall was thinking it could read news links directly. Result: it only returned raw HTML and couldn't even extract tables.

Correct approach: Manually organize data into a table. Zeekr's July deliveries include fields for Model, Delivery Volume, and YoY Growth Rate. I spent 5 minutes creating a table in Excel, looking roughly like this:

Model	Delivery Volume	YoY Growth
Zeekr 9X	8200	35%
Zeekr 8X	7500	28%
Shooting Brake	11000	12%
7X	9137	15%

Note: Data must be clean. No empty cells, no Chinese commas, unified date formats. Initially, I set the number format to "Text," causing Claude Science to throw an error about mismatched data types. It took ten minutes of messing around to figure that out.

Step 2: Upload and Define Analysis Goals

Open Claude Science and select "Data Analysis" mode. After uploading the file, don't rush to ask "analyze this." You must first tell it: What is the context, what do you care about, and what is your analysis goal.

Here is the prompt I wrote:

This data is Zeekr's July 2026 delivery data. I need:

1. Calculate the delivery percentage for each model

2. Determine if growth is primarily driven by a single model

3. Generate a bar chart sorted by delivery volume in descending order

4. Provide a brief conclusion explaining why total delivery grew 111% while individual model growth rates vary significantly

This prompt is critical. If you just write "analyze this," it will automatically attempt clustering, correlation analysis, and prediction, but the results are often irrelevant and slow. Specifying goals keeps it focused.

Step 3: Tuning Parameters and Iteration

Claude Science runs once first, then gives you an "Analysis Report" and code. Note, you need to look at the code, not just the conclusions. The code contains pandas and matplotlib. If you find errors, like wrong colors or axis labels, you can edit directly in the "Code Snapshot" or use the "Regenerate" button.

One pitfall I encountered: It set the bar color for the Shooting Brake model to red, clashing with the blue for the 7X. I added a line to the prompt: "Please use blue-toned bar charts, sorted by model name initial," and it worked this time.

Bold tip: Do not adjust too many parameters at once. Change only one each time. I made three changes: first color, second sorting, third font size. Each time, I submitted, waited for completion, confirmed correctness, then changed the next parameter.

Pitfall Summary: Three Common Beginner Mistakes

1. Non-standard data formats. Claude Science is strict with CSVs; numeric columns must contain only numbers, no "ten thousand" or "%" symbols. Initially, my "8200 units" was recognized as text. Later, I used Excel's "Replace" function to remove all non-numeric characters, and it worked.

2. Vague prompts. Don't write "help me analyze this data." Write "Calculate the proportion for each model, sort them, generate a bar chart, and explain why the total growth is far higher than individual model growth rates." The more specific, the more accurate.

3. Ignoring parameter tuning. Claude Science's default confidence threshold is 0.8, but if your data is noisy, you can lower it to 0.5. I adjusted it twice to filter out an anomaly (Zeekr 9X's 35% growth rate was actually MoM, not YoY; the data was incorrect).

What to Try Next After Learning This

If you can get the above flow working, you've mastered the basics of AI data analysis. Next, I suggest trying:

• Writing paper drafts with Claude Science: Send the analysis results and prompts together, asking it to generate "Methodology" and "Results" sections. You'll find it writes more structurally clear than many grad students, but you still need to manually correct factual errors.
• Combining other data sources: Such as Zeekr's quarterly financial reports or industry average growth rates, for multivariate analysis. Claude Science supports uploading multiple files, allowing you to perform "correlation analysis" or "regression prediction."
• Trying automation: Turn repetitive analyses into templates to run daily. Claude Science's API interface supports programmatic calls, but requires some Python basics.

Trend prediction: Over the next year, AI data analysis tools will shift from "auxiliary" to "core." Those who can use them for exploratory data analysis will lead those still relying on manual Excel work. But the prerequisite is learning to ask the right questions, rather than solely relying on AI-generated conclusions.