Yesterday I caught myself editing a Claude draft and scrubbing the same three words I had scrubbed the day before. Shape. Frame. Lane.
These are not words. Like, something's off with this writing. It's correct, but just, off. Claude reaches for them when it wants to sound clever about a structure. Same shape as the prior idea. A surface is just one place the frame lands. He picked the lane partly for how AI-proof it is. Each one reads as generic. None of them are mine. Every time I make a correction, I tell Claude to log this in a corrections log, which gets updated daily. Cull these freaky word sequences and phrases. I don't ever let Claude just run wild on a post. I nearly always start the post, start with an idea, and then basically riff with Claude as my co-editor, author.
The anti ai corrections log is the durable fat skill that I keep talking about of publishing daily with an AI as co-author. Claude drafts; I edit; the log captures what got cut, what replaced it, and the rule going forward. The blacklist file beside it is longer still. Beautiful applied to a person. Powerful, journey, soul, universe as label words. The darkest stretch as the most common grief phrasing in English. That is the whole trick and here is the part that hurts as announce-the-takeaway sentences. Well, at the start of any sentence. Forced similes that perform wit instead of rendering action: refreshing the applicant list like a teenager checking a crush's Instagram (real example from a Claude draft, scrubbed by me, now banned). The cleaner version is what my human touch and what it actually is. I check the applicant list most mornings. Same empty page.
Claude is fine at writing. The trouble is that fine reads as AI-fine, and AI-fine is the default for everyone now. The work is in the gap between fine and yours. Writing every day isn't easy. Sometimes it's the best part of my day, I wake up and I can't wait to write about the lessons learned from the previous day.
Other days, like today, I'm doing it as the last part of my day. It's a habit I have not broken in 42 days. If I felt myself getting sick, I'd still feel obligated to post something. Maybe about being sick that day and not wanting to write. Why? I know that no one is reading, but because that day, 42 days ago I said I would do something. I'm going to do it.
Some days the edit is twenty minutes, some days it is two hours. Without the log, every post would drift toward the average voice of the internet, which is exactly what models trained on the internet sound like. With it, what lands on this page is mine. Yesterday's trust post made the same argument on the model-trust side. The same discipline applies to voice. Cross-check, push back, keep your own copy. And always, close the loop. If you make an update to the anti ai corrections log, don't fix the article, fix the skill. Then, close the loop.
The model can be typist and the proposer; the writer stays the writer. There is no shortcut around the editing pass. There is only the question of whether the editing pass exists at all.
If you write in public alongside an AI, build the log. The minute you stop catching the tics, your voice is the model's voice, and you have stopped publishing. You have started co-signing, and that's not drift I'm going to allow.
Monthly Revenues $11,800 | Clients 2 | Prospects 1 | Employees: me
Day 42 of 365.