How to write well with AI

Why people who pledge never to write with AI are telling on themselves

I tried to vibecode an app the other day.

It did not go well.

I’m not a coder, so I didn’t really know how to prompt the AI precisely enough to get it to do what I wanted. It spat out a thing that wasn’t really what I wanted, and because I’m not a coder, I didn’t know how to get it to change it into the thing I did want.

Some sort of app did come out the other end. It sucks.

The wrong lesson to take from that experience is “AI is bad at coding.” Software engineers all seem to find that AI is incredibly good at coding. It’s just that to get AI to code well, you sort of have to know how to code yourself. If you heard a software engineer say they think AI is terrible at coding, you’d probably think to yourself “this guy isn’t very good at coding.” And you’d probably be right.

It’s the same thing for writing.

I write with AI all the time. Every day, really.

It’s just that, I’m a writer. I’ve been doing it professionally for 25 years. When AI gives me a garbage output, I can spot where the garbagefulness resides right away. I know exactly what good writing is, so I know precisely how to prompt AI when it messes up.

People who complain AI is bad at writing are sort of telling on themselves. If you’re bad at writing, AI won’t help you much. It’ll clean up your sentences, sure: hunt for typos, give your garbage input some kind of superficial sheen. But it won’t make it good.

If you’re good at writing, AI is hugely liberating. My workflow now consists of a series of very intensive collaborations with Claude. I’ll work up a draft, upload it. Ask it to pick it apart. Consider the feedback, pick apart the good advice from the bad. Rewrite and upload again, ask him to suggest a better way of expressing this point or that, consider it, then tweak it and keep it or chuck it out fearlessly, because the robot never takes an edit personally. I’ll keep iterating, taking advantage of Claude’s true killer feature: indefatigableness. It can keep spotting typos on the 18th iteration, long after any human editor’s eyes would’ve gone dead. I’ll be insanely picky, because the robot doesn’t mind.

Then I’ll ask it to fact check, and to fix any mistakes. Then I’ll ask it to read it oppositionally, looking for weaknesses a critic might attack or imprecisions that might confuse someone. I’ll rewrite again, and check again. On my own this would take an unfeasibly long time, but the robot is fast, and it doesn’t get tired.

And I’ll ask it to help me identify the people I’d need to interview to really make sure it’s right, then I’ll put in a call to that person to check. I don’t expect the robot to replace reporting, I expect it to take the pain out of reporting.

What comes out at the end of this process is unambiguously better than what I could’ve written on my own in a comparable span of time. I might have been able to get to that same level on my own if I had 60 hours to throw at a piece, but I don’t. I have 3 or 4. The end product is fully mine: even if not every word on the page is one I wrote, every word on the page is one I closely scrutinized.

If you’re a writer, Claude can make you a better writer. If you’re not a writer, it can’t do magic. If you have a half-baked idea, prompt the AI sloppily once then copy and paste its first response into the Substack window, you’re debasing the discourse.

But you’re not proving AI is bad at writing. You’re proving you’re bad at writing.