How I used Claude Code in a real data journalism project

This morning three colleagues and I published a story outlining how the federal government is using AI. Here’s how I used Claude Code to help.

Agencies are required to publish a spreadsheet of AI use cases. Unfortunately (and unsurprisingly) each agency posts them in a different place on their website, in a different file format, using different column names and values. OMB will eventually consolidate these, but why wait when we can do it ourselves?

Locating, consolidating and cleaning disparate data sources is a classic data journalism task. It’s also a perfect use-case for “agentic” AI systems like Claude Code. Having done similar projects many times before, I had a good sense of the steps and how to structure the files to ensure auditability and idempotence.

Here’s the prompt I gave to Claude Code (Opus 4.5):

for every agency in data/raw/agencies.txt:

1. search the internet for “AGENCY_NAME AI use inventory". each agency should have an official .gov page outlining ai uses.
2. find in that page the downloadable csv/excel file for the agency use cases. save it in data/raw.
3. when you have found them all, write a python script that converts them all into one csv file. make it easy to read so I can double check that every agency is correct. the columns should all include the same information even if they are not named exactly the same. keep a log of any questions or potentially confusing situations that I should double check.

That churned for about 10 minutes before I hit the usage limit 😭. The prompt fired off a ton of web searches but none completed step 2 by the time I had to quit. I should have had it fill out a spreadsheet as it went. Save all incremental progress to file.

Since I also have a ChatGPT subscription, I switched to Codex and tried just step 1, asking to save progress to a csv. That worked for most agencies. I cleaned that csv up by hand, deleting some results from 2024 and a few other silly errors.

After lunch, I turned back to Claude Code, asking to write a script to download all files from the csv. That worked perfectly. Don’t try to one-shot a complicated process. Go one step at a time.

Next I asked Claude Code to perform the consolidation step. This was the big timesaver. Claude started a loop where it read a few of the raw data files, wrote a script to put them together, ran it for all files, tweaked the script, and on and on. The result is a reasonable python script that would have been horribly tedious to write by hand. I also set up a script to search for new files and download them, and made sure the consolidate script would be rerun without breaking.

To be clear, hell yes I read the generated code. When you’re doing data journalism, vibes are not enough. I have been told “You’re absolutely right!” far too many times by these tools to trust them. Likewise I would not trust an LLM to read the data directly. But having AI write and execute code that can be audited? I’m quite comfortable with that.

After a lot more spot checking, it was ready to share with the team for further analysis. Not bad for a few day’s work.