I Was Iterating On The Wrong Layer

Until a few weeks ago, the way I wrote this newsletter contradicted what I’d been telling my readers to do.

Every issue is about using AI well without burning out. And every issue, I was burning out to write it.

My first issue went through sixteen drafts before I sent it. The next few weren't much better. I was spending six to eight hours writing each article. Most of it spent fighting whatever Claude gave me.

The first thing I'd do was try to edit Claude's draft. I'd move a sentence, tighten an opener, cut a paragraph. Eventually I'd give up on word edits, write a long list of notes about what was wrong, and ask Claude to rewrite the whole thing. Then I'd have to read it carefully to see whether Claude had applied the notes. By the time I figured out which parts had landed and which were still off, I'd lost another hour.

Claude can write a 1,500-word article in less than a minute. But it could take you over an hour to properly read, comprehend, and determine if it’s what you actually want to say. What you initially get back feels 80% done. The 20% that separates real writing from slop is where the hours go. The output is free. The judgment is not.

The real problem was never in any single output. I kept fixing the current article instead of fixing the way Claude and I built articles.

I was iterating on the wrong layer.

I couldn't “judge” any faster than I already was. But what I could do was make my judgments stick. So I started telling Claude not just what I wanted fixed in the current article, but what I wanted to improve about how we worked together. It was a small change with such a big impact that I'm still kicking myself for not doing it sooner.

The first time was almost by accident. I'd just spent an hour cleaning up a draft that opened with three lines of a fabricated story. Out of desperation, I told Claude what had gone wrong and asked it to never do it again: don't make up stories to fit the narrative!

Claude saved my desperate pleas in its memory. The next article, the mistakes I complained about didn't happen. But a different one did. So I added that rule. Then another. Then another. After just a few weeks, Claude had already collected forty-eight specific writing rules, plus a couple dozen more rules about how we should work together.

A little extra work at the end of each session, and each subsequent article got easier. An article that used to take six hours now takes three.

If you've ever had this problem yourself, here's what to try this week.

1. Stop iterating on output. Iterate on the process.

When your AI gives you something you don't like, don't regenerate. Pause. Name the specific reason it's wrong: the tone, the example, the structure, an invented detail. Tell your AI the reason and ask it to remember the rule for next time. Each rule you add gets your AI closer to what you actually want.

2. Start with many small checkpoints. Earn the right to skip them later.

Start with the big picture, then check in at each step before going deeper. For most work, that means locking the goal first, then the structure, then the bullets for each section, before asking for any prose. The shape of "the small thing" is yours to design.

Expect the early sessions to be slow. That slowness is where the structure comes from. After a few iterations, you'll spend less total time than you do now.

3. Make every rule persistent.

It's not enough to add rules during a session. Save the ones that matter where your AI will read them next time. A file in your project. The top of your standard prompt. A "guidelines" doc you paste into every new chat. Whatever your AI looks at first when you start working.

A rule itself can be a single line of plain language. Things like "Show me your reasoning, not just the answer." Or "Wait for my approval before scheduling anything." No special format. The AI reads it and applies it.

Skipping this step is why most setups never compound: every session starts cold, and the lessons from the last one stay in the last one. The skill is making this week's lessons available to next week, and the week after.

I've been writing this newsletter for almost ten weeks, and the concept of memory and iterating on learning has been touched upon in half my issues. I feel silly for taking so long to implement it in my own writing, but better late than not at all.

me+machine