The Rise, Fall, and Imminent Return of the Em Dash

A few friends have started using AI more regularly. I'll take partial credit for that, even if it's coincidental. But watching how they're using it has been, to be honest, a little uncomfortable. The writing is cluttered with em dashes. The syntax is off. You can tell immediately that nobody cleaned it up before hitting send. I'm not judging the impulse. I'm judging the finish.

The issue isn't laziness. I don't think people are cutting corners on purpose. The issue is that we've imported the wrong mental model from every other piece of software we've ever used.

With traditional SaaS, the logic is simple: you put something in, you get something out, you're done. The platform just works. That instinct made sense for decades, and every software company in the world is working overtime to rebrand themselves as "AI-enabled" precisely because it no longer does. But the reflex hasn't caught up yet.

There's a meme circulating right now of people prompting Claude to build multi-billion dollar tech companies in one shot. "Rebuild Uber from scratch. Make no mistakes." It captures the instinct perfectly.

AI isn't like that. The first output is a starting point, not a deliverable. When you prompt once, accept the result, and send, what you get is AI slop. And AI slop has tells. The most obvious one right now is the em dash.

The em dash itself is fine. It's a useful punctuation mark. You'll notice I'm not using any in this issue. That was intentional. Mostly. But AI defaults to them constantly, and when your writing is covered in them, people discount what you've written. They may not say anything. They'll just quietly apply less weight to your work. Over time, that adds up.

Most people won't catch this themselves. They'll get corrected by a boss, a client, someone whose opinion they care about. The feedback will probably be gentle: "I'm glad you're using these tools, but can you tighten it up a little." That's the version you want. The worse version is they stop taking your writing seriously and never tell you why.

This tweet is emblematic of the moment. The author's name is actually Claude, which I find funny and appropriate. The early adopters are reclaiming the em dash. The mass adoption wave is still in the slop phase. The retired professional who just found Claude last month and is sending emails covered in dashes to people who notice these things will find out eventually.

My prediction: the em dash survives. It just has to earn its place again.

I want to introduce you to something that will meaningfully improve your AI output. Not a magic fix, but a structural one that forces you to do the work upfront that most people skip entirely.

Projects are dedicated workspaces inside Claude and ChatGPT. Instead of starting a new chat from scratch every time, a Project gives Claude persistent context: who you are, what you're working on, how you think, what you want. You set it once. Claude carries it forward across every conversation in that project.

A Project has two components that do the heavy lifting.

Project instructions. This is where you tell Claude what it needs to know to be genuinely useful. Your role, your goals, your preferences, your audience. Think of it less like a bio and more like a brief. The more specific, the better.

Project files. Documents, reference material, past work, anything Claude should be able to draw on. A lawyer might upload a client summary. An executive might upload their team's strategy doc. The point is that Claude is working with your actual context, not guessing.

The instinct most people will have when setting up a Project is to write the instructions themselves. A few sentences, maybe some bullets. Resist that. The better approach is to ask Claude to interview you. Tell it you're setting up a new Project and want help writing the instructions. Let it ask questions. You'll end up with something far more complete than anything you'd have written on your own, and the process surfaces things you didn't realize Claude needed to know.

On hygiene: monthly is probably right for most people. If you're rewriting your instructions every week, the Project scope is too broad or you're pivoting too often on the underlying work. The instructions should have some endurance. Think of them as a standing brief, not a to-do list.

I work in Claude and recommend it. One reason among several:

A well-maintained Project is the difference between a conversation with someone who knows you and a cold call. The output reflects that immediately.

Create a Project in Claude or ChatGPT this week. Don't write the instructions yourself. Open the Project, tell Claude you want help setting it up, and let it interview you first. Add at least one relevant file. That's it.

New York is Claude's biggest market. Relative to population, New York uses Claude more than any other U.S. state, trailing only Washington, D.C., per the NYC Comptroller's February 2026 jobs report. My read: New York is the capital of capital. If Anthropic is building a platform for serious professional work, that's exactly where you'd expect the heaviest adoption. My other hypothesis: Anthropic tends to align with progressive values, and New York is about as blue as it gets.

A man used AI to build a cancer vaccine for his dog. An Australian tech entrepreneur used AI to help develop what may be the first bespoke cancer vaccine for a dog, his rescue Rosie, working around regulatory channels to do it. I want to be careful: I'm not scientifically qualified to call this a victory, and the jury is still out on whether the vaccine was responsible for Rosie's improvement. But I love this story. Amid all the noise about what AI might take from us, here's a case of a guy using it to fight for his dog. On a personal note, I have a rescue too, so this story hit home.