Every morning I read AI news for about 10 minutes on Nitu AI with my coffee, and by the end of the week I always have something worth writing about. What I want to show you today is not the reading part. It's the system behind my newsletter, the one that does the research, drafts the issue, fact-checks it, generates the images, and gets it ready to publish, while I still read and approve every single line before it goes out.

Why I didn't go full zero-touch

I could have built this so the newsletter publishes itself with no review. I didn't. I still open the Word document, read it, and change what I want to change before anything goes live.

And that's on purpose. I tested letting AI handle everything end to end for other projects, and the pattern I keep running into is simple: readers subscribe for your voice, your opinion, your take. Not a generic summary anyone could have written.

There's a real cost to skipping the review step too. I know of a case where a financial newsletter used AI to summarize a regulatory change, got the date wrong, and subscribers who acted on that wrong date lost money. That newsletter lost 30 percent of its list in one week. That's not a small mistake, that's the whole business taking a hit.

So my rule is simple. AI does the research, the drafting, the fact-checking, the image generation, the formatting. I do the final read.

How the research and writing actually works

I built this inside a Claude project. All I do is paste a link or an idea, in this case a story about ChatGPT ads expanding across Europe, and tell it this is the next issue.

Claude doesn't just take my link at face value. It searches the web to confirm other credible sources are reporting the same thing before it writes a single word. That's a rule I built into the system on purpose, because I don't want to publish something based on one shaky source.

Once it confirms the story checks out, it asks me a set of questions I set up ahead of time. What's the angle or goal for this issue? Who's the audience? Do I have a personal story to weave in? What opinion should the piece take?

For the ChatGPT ads story, I picked the emotional bond angle, since people talk to ChatGPT like a confidant, not like a search bar, so ads sitting inside that kind of exchange feel different from a banner ad. I said no personal story for this one, and I asked it to present both sides neutrally rather than picking a side outright.

After I answer those questions, Claude writes the whole thing: LinkedIn blurb, hashtags, image prompts, a TL:DR section, six sections of the newsletter, my sign-off, and sources at the bottom for credibility.

Generating the images for under 10 cents

Once I approve the written draft, Claude writes image prompts and sends them straight to the OpenAI API. Before generating, I had $5.91 in credits sitting there.

I needed six images for this issue. After generation finished, I checked the balance again and the whole batch cost less than 10 cents. That's genuinely cheap for six on-brand illustrations.

And they're actually on-brand, not generic. The prompts pull from my own color palette, so a story about ChatGPT ads turned into an image showing a conversation that used to be clean now filled with ads and price tags, in my exact navy and blue tones. That's a small detail but it means I'm not manually matching colors in Canva afterward.

From Word document to published newsletter and email

I always get a Word document out of this, and that matters to me because if there's anything I don't like, I can still edit it before it touches my website.

Once I'm happy with the read, I run a script that asks me where the newsletter folder is and what publish date I want. I usually publish weekly on a Wednesday, so I checked my last issue went out August 26 and set this one for September 3, 2026.

The script also compresses every image so none of them go above 80 KB. I noticed some of my raw images were sitting at 1.24 megabytes each, and if you're publishing five or six of those every week without compression, your site gets heavy fast.

Once compression is done, I click approve, the issue publishes to my website, and a separate email file gets generated. I copy that into Kit, my email provider, using an HTML block inside a template I already built with my brand colors, and it's ready to send.

Keeping it from sounding like AI wrote it

If you want to build something like this yourself instead of having someone build it for you, the writing quality is where most of these systems fall apart. Readers can tell when copy is generic.

I feed Claude two documents to fix that. One is an anti-AI writing skill file that trains it to avoid the stiff, over-formal patterns AI text tends to fall into. The other is a humanization skill built by Sabrina Ramonov, whose work on this I genuinely rate.

Both are free, and I include them for anyone signing up to my Boundless Automation newsletter. Check the welcome email for the resource link.

One more thing worth flagging if you're sending through an automated agent instead of a human clicking send: agents can fire off newsletters in sudden bursts with no natural rhythm, and that pattern can trip volume-based spam filters, since it looks less like a human sending at irregular intervals and more like bot behavior.