I Do My Best AI Work After I Shut the Laptop

I do my best AI work after I shut the laptop.

For agency, judgement and taste.

Because as code is cheap — “Pure software is rapidly becoming uninvestable.”

Disconnect to keep agency and judgement ahead of AI output

I can spin up ten Claude agents in a few minutes. And cross-prompt with GPT-5.5 Pro and Gemini. They write more than I can possibly judge.

Speed is no longer the bottleneck. Agency is.

The only time I have real agency is when I'm away from my screen.

The /remarkable loop

If I don't pause, I spin up another agent whilst waiting and keep going!

When a strategy, design or plan needs real thought, I run /remarkable.

It turns the markdown into a proper PDF — large type, generous margins — and uploads it to my reMarkable.

If you don't know a reMarkable, the core features I love:

• Write on it and it feels like paper.
• No internet — pure disconnect and focus.

I go to the beach. Read slowly. Push back on AI in the margins. Come back with the file, my notes, and a clearer opinion.

The next AI round is better because I'm not asking it to think for me. I'm asking it to respond to what I actually think.

Without that hour I'm just running an agent farm.

(The backbone of /remarkable is the open-source rmapi, maintained by ddvk. Thank you.)

What I'm reading

Naval Ravikant — A Return to Code. A new essay he published last week. The line that opens this email — Pure software is rapidly becoming uninvestable — is his.

But his real point is:

“There's never been a better time to be alive as a creator of software.”

And that's what I feel.

I'm operating on a whole new level and loving it.

Marc Andreessen — The Real AI Boom Hasn't Even Begun, Lenny's Podcast (yes, I'm reading YouTube videos on my reMarkable now — more in a moment):

“The additive effect of being good at two things is more than double. The additive effect of being good at three things is more than triple.”

I am feeling this with AI coding. I'm compounding skills of design + more + and now code, and outputting things I couldn't dream of a year ago.

Rahul Vohra (Superhuman) — How Superhuman Built an Engine to Find Product-Market Fit. The Sean Ellis test in plain English: ask your users how they'd feel without your product. 40% “very disappointed” is the bar. Why I'm running it on Design Canvas (more on that below).

The /youtube script

YouTube has quietly become the best place for current thinking.

I know — I've said before, no social. But YouTube feels different. A source of information and learning, not ego or algorithms designed to make you feel inadequate. (If you can ignore the fake MRR videos.)

Watching it is the problem.

YouTube video = hook → watch for 40 minutes → find a snippet of useful information at the end.

So I built a script.

I email a YouTube link to a private address. Claude pulls the transcript, reads it against my own planning files, and ships a personalised PDF to my reMarkable.

This email used the same loop. Drafted on the laptop. Read on the beach. Argued with in the margin. Sent.

Last email I shipped Design Canvas. Since then — five user tests. Szymon (Ramp.Network Founder) told me “I need this to exist.”

Hopefully we're on the route to PMF.

Two iterations later. Latest is live.

Life

Lots of hiking with Coco and Bene.

Also very needed to switch off from all the knowledge, designing and coding for clients and my own apps.

If there's something interesting in here, I'd love to know what it is — or your own experiences. Also ping me if any of these scripts are useful and I can share. Although armed with Claude Code, you can describe one in a prompt and have your own. I love getting email.

This is email four. Previous: ellington.design/emails. For the AI co-founder thread, see email two.