Howdy wizards,
When you’ve decided to automate a process with AI, it’s easy to think too small.
I’ll show you an example.
You might’ve noticed the cover image of this newsletter has a certain style to it.
That bearded character (the Wizard), always with a coffee cup, always working on his laptop. He’s the face of this brand.
Yeah, he’s AI generated. But I have my own prompt and reference images I use to make the character and environment look consistent.
My point is this: I used to spend 20 minutes on this. Every week.
What making one cover image involved:
Prompting
Choosing between images
Placing the images into my Figma template
Exporting it, compressing it and putting it into my newsletter
In my head, I naively thought I was already using AI to its full potential in this process.
After all, I was prompting.
But I was wrong.
I was using AI for only a tiny part of the actual process.
And each week I’d drop what I was doing to sit and generate images.
I’m at edition #237 of this newsletter. That’s a full work week, give or take, that I’ve spent generating cover images!
Not exactly a good ROI situation.
Eventually I realised I could pretty much use AI for the whole thing.
I had a mindset shift.
That shift has helped me automate many boring processes in my work since then.
I’ll show you what the mindset looks like, and what my new cover image workflow looks like, too.
One thing I still haven’t used AI for is running a voice agent.
If I did, I’d have a caller’s voice mixed with everyone else talking in the background. Hello messy transcript, and an agent taking all kinds of wrong actions. I’m glad that’s not my problem. But if it is yours, definitely check out this week’s sponsor, Krisp:
IN PARTNERSHIP WITH KRISP
AI agents that answer phone calls work great in demos, then fall apart on real calls. The reason is usually the audio.
Speech-to-text nowadays handles noise just fine. Traffic, fans, air conditioning, no problem. What breaks it is a second person talking near the caller. A noise filter can’t tell one voice from another. The transcript comes out wrong, and the agent acts on it: wrong answer, wrong booking, “could you repeat that?” until the caller hangs up.
Krisp’s Voice Isolation 2.5 sits in front of your speech-to-text and isolates the caller’s voice from everyone talking around them. On calls with a competing speaker, word errors drop from 36% to 11%. It adds 15 ms of latency and barely touches clean audio, so you can leave it on for every call.
❦
The mindset shift
Here’s how I think about using AI at work nowadays:
Direction & taste are your human currency. The rest can be delegated to AI.
It feels scary to take on this mental model.
Like — am I gonna lose every last brain cell now?
But I already did that. I asked AI for everything and lost all my brain cells.
This is different. I don’t ask AI for everything, I use it for everything.
I go to AI for execution, not answers.
I go to AI when I know what I want, and why I want it.
Solving the technicalities of how to get there is mostly AI’s job. My role is to get out of its way.
Most of the sub-processes of any task are possible for AI to do.
It will require your deep cognitive involvement in the process. You will have to give AI access to things. You will have to test and you will have to verify. There will be trial and error.
But if you keep at it, you’ll get there. And over time, you’ll understand the common obstacles AI faces, and be able to do it more and more effectively.
Enough theory.
Here’s how I solved the cover image workflow.
My first step was to map out the actual steps of the process:
What I called it | What it actually was |
|---|---|
Prompting | Finding the prompt in my notes app. |
Choosing between images | Comparing five tabs. Downloading two. Discovering both look wrong once they’re in the frame. Going back for more. |
Placing them into Figma | Opening the file. Finding the right element. Dropping it in. Nudging it until the crop stops eating his face. Checking it against the grid so three issues in a row don’t look identical. |
Exporting and putting it in the newsletter | Exporting. Too heavy. Uploading it to some compression website. Downloading it again. Creating the beehiiv draft. Dragging it in. Checking the edition number. Getting the edition number wrong. |
I gave Claude Code what’s in the table above and said:
“Help!”
“Build a workflow for creating cover images that look on-brand every single time. But not the same. Him in a café, on a train, in a car. Coffee in hand, coffee on the table, coffee somewhere. Laptop open, laptop closed, laptop under his arm. I want all of that, and I want to control it by dragging a slider.”
“I want this fully automated so all I need to do is sit back, look at nice cover images that have been prepared for me and pick one.”
After some back and forth, Claude came up with the plan. It consisted of two parts:
Building an app:
A gallery where I can see the cover images of previous and upcoming editions in a grid
A built-in image editor where I can select between pre-generated cover images for each edition, and make edits to them just using sliders and dropdowns
Some background jobs:
Using Gemini API to generate the set of images for each edition according to that edition’s specs. I can easily generate more variations in the UI.
Communicating with my email platform (beehiiv), and inserting the correct cover image into the correct newsletter edition automatically.
An hour later I had a working app. Then a couple of hours more of back and forths.
I have full overview of previous and upcoming cover images in a grid
I can select between pre-generated images for each newsletter edition, generate more, and edit any aspect of the images easily
The only manual job I do now is look at pretty images and select the best one.
By the time I start writing my newsletter, the cover image is already in there.
Occasionally I use the option to change the camera angle, his location, posture, colours, etc.
I have the option to de-select the coffee cup, as a joke.
He will never stop drinking coffee.
What’s left of the process is the part I actually enjoy. Looking strong options and picking the best one. I can’t put into words exactly what I want a cover to be, but I know it when I see it.
My job now is taste.
Realising AI can do most parts of anything is scary.
You’re allowed to sit and be scared.
You’re also allowed to pour yourself a fine cup of coffee, get a Claude subscription and lean into it.
(And don’t forget, a good playlist.)
❦
You are a delight.
What's your verdict on today's email?
See which AI use cases are paying off with Context Windows Pro
Most companies pick AI use cases by brainstorming internally. 90% of those initiatives fail.
I’ve created Context Windows just so you can pick the winners.
🟦 Find high-performing use cases from 2,000+ companies at contextwindows.ai, or book a demo with me
Disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes work with sponsors and may earn a commission if you buy something through a link in here. If you choose to click, subscribe, or buy through any of them, THANK YOU – it will make it possible for me to continue to do this.