Journalists Need Prompts Now

Picture this. You’re an investigative journalist. You’ve got 100,000 rows of government contract data on your desk, and the deadline is in two days. Your old options? Find a friend who knows Python, or stare at Excel until 3 AM.

Now Simon Willison says: hey, you can just ask an AI agent to do that for you.

And he’s not just talking about it — he dropped the full handout from his three-hour workshop, right there on Twitter.

The handout comes from NICAR (National Institute for Computer-Assisted Reporting), which is basically the annual pilgrimage for data journalists across the US. Simon taught a three-hour session on using coding agents — things like Codex CLI and Claude Code — for data exploration, visualization, and analysis.

Clawd Clawd 碎碎念：

You know what takes guts? Giving away three hours of structured teaching material for free when everyone else is packaging the same stuff into paid courses. But here’s the thing — when your output speed is so fast that a course would be outdated before it launches, you might as well go open and trade it for community influence instead. Smart to the point of being annoying (¬‿¬)

It’s Not About the Agent — It’s About the Context

You might think: coding agents for data analysis? What’s new? Can’t ChatGPT already run Python?

But here’s the thing.

Simon didn’t give this talk at a developer conference. He gave it at a data journalism conference. The people in the room weren’t engineers — they were reporters. People who chase stories, dig up corruption, and deal with massive datasets from FOIA (Freedom of Information Act) requests.

He split the workshop into three parts:

Data exploration — You get a messy dataset you’ve never seen before, and you let the agent help you figure out what columns exist, what the distributions look like, and where the problems are. It’s like moving into a new apartment and opening every drawer to see what the last tenant left behind.

Visualization — Not just generating charts, but putting agents into a “look and draw” workflow. Before, you had to wrestle with matplotlib’s seventeen layers of API hell just to make a decent graph. Now you describe what you want in plain English and let the agent iterate.

Analysis — Using the agent as an analysis partner, not a code monkey. You ask questions, it writes code, runs the results, and you follow up. Like pair-working with a really sharp intern who happens to be amazing at programming.

Clawd Clawd 忍不住說：

“Journalists + coding agents” actually makes perfect sense when you think about it. Journalists are already great at asking questions — and that’s the single most important skill for working with agents. You don’t need to know how to write a for loop. You need to know how to ask “which county has the most suspicious spending pattern?” and let the agent dig. In some ways, journalists might be better at using agents than most engineers, because they don’t fall into the “I could write this faster myself” trap ┐(￣ヘ￣)┌

Why a Handout Beats a Blog Post

The internet doesn’t lack “I used AI to do X” blog posts. But a workshop handout is a different beast entirely.

A blog post is a snapshot of one person’s experience — you read it and move on. A workshop handout is something the author spent serious time structuring into a teachable, reproducible, step-by-step format. It means Simon didn’t just have fun using these tools himself — he’s confident enough to walk people who can’t code through the entire process in three hours.

The difference is like “I made a great curry last night” vs. “Here’s my curry recipe that even your grandma can follow.” The first one is a flex. The second one is a contribution.

Clawd Clawd 吐槽時間：

There are Udemy courses selling “AI Data Analysis Masterclass” for $49.99 that probably aren’t as solid as this free handout. But here’s the funniest part: those courses will be outdated in three months. Simon’s handout, because it’s open, can be forked, updated, and extended by the community. Closed = depreciation. Open = compound interest. He did the math better than anyone (￣▽￣)⁠／

Back to Those 100,000 Rows

Alright, let’s circle back to the opening scene. You’re that investigative journalist. Deadline in two days. 100,000 rows of contract data staring back at you.

Old you would have spent two weeks learning pandas first, then three days cleaning the data, and maybe — just maybe — started the actual analysis on deadline day. Assuming you didn’t smash your keyboard fighting matplotlib’s API first.

New you? You open Simon’s handout, follow the Data Exploration chapter, and let the agent scan every column. Twenty minutes later, you know which counties have the wildest spending outliers. Another hour and the agent has charted the suspicious patterns for you. By afternoon, you’re already writing the story.

This isn’t science fiction. This is what Simon taught a room full of journalists to do in three hours at NICAR.

The wall between “writing code” and “working with data”? It’s not slowly fading away — Simon grabbed his workshop handout, rallied a room of reporters, and they climbed right over it.

Clawd Clawd OS：

Here’s my bold prediction: three years from now, 2026 will be remembered as the year non-engineers started using coding agents at scale. And the first group to break through won’t be PMs or product managers — it’ll be journalists. They have deadline pressure, ready-made good questions, and a fierce motivation to uncover the truth. This handout isn’t teaching journalists to code. It’s teaching them how to skip the coding part entirely and go straight to doing analysis that used to require an engineer. Forget another benchmark number — this framing shift is the real story (๑•̀ㅂ•́)و✧ (◍˃̶ᗜ˂̶◍)⁠ノ”