AI moves fast. Some weeks there are 80 relevant news pieces to get through. Most weeks it's around 40. It was taking me nine hours a week to read, digest, sort and write the whole thing up. Add in being dyslexic - it was all a bit of a nightmare.

As I'm actually busy running a business, I tried using AI to speed things up. Here's a guide so you can learn from my trial and error.

What Didn't Work (And Why)

First attempt: ask ChatGPT to research and write the newsletter. This ended up taking me more time. It was awful. It made things up - links that didn't exist, stories that hadn't happened. It kept including things we'd already covered, drifted off topic, and didn't really have any sense of what had happened specifically that week. Garbage in, garbage out (to use the polite version).

Second attempt: give ChatGPT the links from me reading the news. This basically just fixed spelling mistakes and split the stories into our three sections. Definitely sped things up, but I kept hitting memory issues.

Projects: setting up a ChatGPT project helped a lot. A project, for those who don't know, is like a workspace - it has instructions and files that the AI can always access, and you can start multiple chats within it. But it only worked reliably with GPT-4o (which is now gone), and even then it was hit and miss.

So we moved to Claude. Noticeably better at staying grounded in what we'd actually sent it. The writing was closer to our tone. The one slightly annoying thing is that Claude occasionally decides something isn't newsworthy and quietly leaves it out - sometimes fair, sometimes frustrating. But it always flags it, so it can be corrected.

One other thing Claude does that ChatGPT didn't: it reorders the links by relevance rather than just listing them in the order I sent. And it adds useful subheadings when there's a lot of news in one category. We do sometimes disagree on what counts as news versus not quite news - but that's an editorial conversation, not a problem.

Getting The Tone Right

This bit took some iteration. I didn't just give Claude a style guide - I loaded in past newsletters and some of our blog posts, and then asked Claude to summarise our tone of voice and style into a prompt. Essentially: describe how we write so I can give that description back to you when writing new issues.

That prompt has evolved over time, but the core of it is:

Write a weekly AI newsletter for Taught by Humans. Use British English. The tone should be friendly, clear, and a little sceptical - curious, but never dramatic, salesy, or sarcastic. It's okay to have a quiet opinion, but don't force one for every story. Prioritise clarity and usefulness over hype.

There's also a line in there that I'd say is the most important thing in the whole setup:

Never rewrite. Only make suggestions.

Some sections of the newsletter I write myself - our week at Taught by Humans, the poll results, the wrap-up. I literally put [I WRITE THIS] in caps in the instructions so Claude doesn't touch those sections. It works. The instruction to only suggest (rather than rewrite) means I stay in control of the editorial decisions. It feels collaborative rather than like I'm just cleaning up AI output.

The Actual Process, Step By Step

Throughout the week: the whole team reads AI news as it comes in and drops links into a Slack channel. By Tuesday, we have a running list of everything worth considering, with links attached. Something like:

Goldman Sachs taps Anthropic's Claude to automate accounting and compliance roles - [link]
Free Claude just got an upgrade - longer memory, file creation - [link]

This is the raw material. It also means we're actually reading the news, which matters. Staying up to date is part of why we do the newsletter.

On Tuesday: I paste the week's links into Claude, it drafts the newsletter sections, and I edit. Before the full draft, I ask for five title and subtitle options - I'll usually like the title from one and the subtitle from another, say so, and iterate from there. A couple of rounds and we have something worth building on.

Then: our written sections go in, final read, copy to Brevo and send. The last step is still manual. We haven't automated it yet.

What We Could Do But Haven't

We could connect the Slack channel directly to Claude. We could probably automate the Brevo step. We could build something that surfaces relevant news automatically rather than relying on the team to spot it.

We haven't, partly because the current process works, and partly because we think it matters that we actually read the news ourselves. It keeps us sharp. When we talk to customers about AI, or write about it, we know what's actually going on - not just what an algorithm decided was relevant.

The automation is there if we need it. For now, the human bit is intentional.

The Actual Time Saving

Nine hours a week down to around 20 minutes to an hour on a Tuesday, with the news trickling in throughout the week rather than all landing at once.

If you run a newsletter, send regular updates to your team, or produce any kind of recurring written content, the same approach works. Build the context once, get the tone right, keep the human bits human.

We use Claude for parts of our newsletter. We also build learning tools on top of Claude and other AI models at Taught by Humans. If you're thinking about how AI can work in your team's day-to-day - get in touch.

How We Actually Use Claude to Write Our AI newsletter (a Working Process, Not a Magic Fix)

Enjoyed this blog?