We're not going to talk about AI

Hey Reader,

First, some literal housekeeping — you may have noticed that, on your phones or other narrow viewports, these emails had like no margins. Even if they were still readable, they looked ugly, and you would be forgiven for thinking, "who is this guy to tell me what good design is or what AI means? His emails are crap!"

Well, for this issue I have fixed the issue. Please enjoy 20 pixels of space between these words and the edge of your phone.

Last weekend I had the pleasure of guesting on Chris Enns' Dailyish — a show he's doing daily as a challenge for his 50th year on this planet — where we covered topics like AI, the excellent Apple TV documentary series Long Way Up, AI, the new IKEA-produced Skyrim mod starring Matt Berry, my expensive Teenage Engineering podcast mic, and I think also some AI.

​

In addition to being on this podcast, I spent a lot of time this week on Content Marketing™, writing some content designed for search engines and AI agents to read and put me into their indices as an expert in where B2B marketers should go after migrating away from Webflow.

Writing this post (which is slated to go live next week) had me thinking about the aesthetics and nature of slop, and whether or how much it matters to make sure text is human-crafted when its #1 consumer is a bot or agent. Listicle or guide-type posts like this tend to follow a rigid structure, and even before AI it was hard to see any variation or spark of creativity in them.

This is meant only somewhat as a dig. In the same way that a weather report is going to feel (perhaps ironically) drier and more formulaic than a human interest article, sometimes prose has a job to do that supersedes whether it sounds authored or not.

However, thanks to AI, we live in an era where formulaic writing can be misinterpreted as having been generated, not written.

You could argue that SEO-bait listicles were always more generated than written, and you probably should never have trusted that a list of "Top 10 alternatives to [thing that competes with PostHog]" that was written by PostHog, where PostHog is the number one alternative to consider, was probably not a well-reasoned opinion piece.

But now we have AI detectors. We have them metaphorically, in that we've trained our brains to look for em dashes and, upon finding one, stop reading, block the sender, raze their city to the ground and salt the earth so no slop will again grow there. We also have them as handy Chrome extensions — Pangram, who make a market-leading AI detector, have one that will add badges to LinkedIn or X posts telling you whether they were partly or wholly AI generated. These pair well with the new-ish LinkedIn feature where you can report posts as "Seems like AI slop."

I discovered something surprising while working on my SEO-slop blog post: while AI can't write well when starting from a blank page, and often can't edit well either no matter what skills you download, the latest frontier models are not bad at using your own line edits in a document to learn how to line-edit other parts of the same document.

I've become a big fan of Claude Docs, the new word processor built into Claude. A Claude Doc is more minimal and focused than a Google Doc (and may even be plain Markdown behind the scenes), and you can tag Claude in comments or new lines to ask it to generate or revise text. The AI can see all your edits in the doc's version history, and from that can develop a much more accurate model of what you'd want to change in your writing.

I had Claude extract what it learned from my edits into a skill, then tried to use the skill both to write fresh copy or to edit an entire draft end to end. Both efforts sucked badly.

But if you ask Claude (or, presumably, GPT Sol or Astra) to use your edits to revise just one paragraph at a time, and you already manually revised a different paragraph in the same doc, it actually does a good job!

Even though my forthcoming Ultimate Guide to Webflow Alternatives post was largely AI-edited, none of it was AI-written, and thanks to my trick of teaching the model my ways one paragraph at a time, the whole thing scores 100% human on Pangram.

I still wonder: does this even matter? Is there a human reader who'd open a post titled "The best Webflow alternatives for B2B marketers" and expect to see human craft at work? 🤷

​As I wrote in an X thread, though it feels like I'm polishing and gold-plating text that's mainly a sales tool, and this genre of post doesn't need that, I'm the human available to do the job so this is what the post is gonna get.

New on the B&L blog

​Agents need a schema, too — Another thing I'm working on presently is sharpening and narrowing how I talk about B&L — what the marketing gurus call positioning — from "websites" to content systems. (Which happen to include websites, but are distinct from the Webflow thing your cousin made both in having an enterprise-grade CMS and costing an order of magnitude more money.)

In this post, I connect the AI stuff I've been blogging and newslettering about with my company's main job, recapping what a content model (or schema) is, why unmodeled content like WordPress's slop-filled post_body fields aren't ideal, and how agents work with all of this.

Lest you think this isn't a real topic, that I'm just looking for an excuse to keep saying "agents" over and over, B2B marketers are indeed keenly interested in how agents can help them automate working on their websites. In the most extreme cases, they're looking to replace agencies or in-house teams entirely with AI tools. But I've had several projects this year (plus a bunch of sales calls) where part of the brief was to not just build a new site, but set it up so the team could handle some ongoing maintenance and change using AI.

For example, back in June we launched a new website for Rill, a data visualization startup. At launch, the new site consisted of a few pages/sections powered by Astro and Sanity, with other "long tail" site pages falling back to their old site on Webflow. In the 2-3 months since launch, the small team was able to finish the migration, with their agents and human marketers keeping things fast and gorgeous using the design system and other rails B&L built in the initial engagement.

More links worth checking out

🏇 Fast is a design decision — In this post by Shanna Dige, "fast" means both in the browser (loading times, compute, Core Web Vitals) and in the process of building (these days, of course, using AI). One AI fallacy that's, thank god, seeming to fall aside as we approach the end of this year is that, because building is now cheap, shipping good products and experiences is also cheap. This post is a good illustration of how serving users well still takes time and attention — the parts of the process that always felt slowest are the most valuable, and at best automation can just reduce the hands-on time needed to implement, test, and analyze the results of our design conversations.

💥 Fail small to reach further — My friend and fellow Xoogler Alex Komoroske, now founder & CEO of Common Fabric, has been banging various drums about the need to improve how AI and humans interact, and long before that was notorious within Google as one of the sharpest thinkers on how systems and people come together to shape the world. (This is his legendary "go/slimemold" deck, adapted for civilian reading.) Here, Alex argues that AI might safely and securely scale to do all the things if we place it into ever smaller sandboxes, shaping our 'trust model' for agents along similar lines as what David Weinberger wrote about the Web more than 20 years in his book Small Pieces Loosely Joined.

Heeeeeere's Johnny

He doesn't spend all his time on the couch, just most of his time. This photo is a rare glimpse of him putting his paws by his cheeks, like an adorable mouse that happens to fart a lot.