AI runs my marketing. I run the AI.
Hey,
Every week this business publishes four LinkedIn posts, a daily drip on X, two emails, and blog posts. AI drafts most of it, and almost none of my time goes into typing first drafts anymore.
That’s the boring part. Everyone’s doing some version of that, and most of it reads exactly like what it is.
The difference is everything I built around the writing.
It remembers every correction. When I fix something (“stop using that phrase”, “never open a post that way”, “that claim needs a source”), the correction gets written down as a rule, and every future draft is checked against the full list. That list is long now. A generic AI tool starts from zero every session, but mine starts from every mistake it’s already made.
It’s not allowed to invent anything. Hard gate: no made-up stats, no invented client stories, no “roughly 30” when the real answer is “I don’t know.” A draft that can’t trace its claims gets thrown out, not polished. This one rule does more for sounding human than any prompt about tone.
Nothing ships raw. Every draft passes through an editor layer with its own checklist: my banned-phrase list, structure patterns that scream AI, fact checks. Then I read it.
The last pass is me, every time. After the editor layer, I go over the draft and make it my own. I cut lines I’d never say, add what actually happened this week, rewrite until it sounds like me talking. Sometimes that’s two minutes, sometimes I rewrite half of it. If I can’t make a draft mine, it doesn’t go out. You’re never reading raw machine output here.
The data feeds back in. Post metrics get pulled into a database automatically every week. A year of posts, scored by hook style, topic, day of the week. New drafts get compared against what actually worked.
The automations watch each other. This week one of my watchdog jobs pinged me because I skipped a weekly step. I built automations to catch me, not just to replace me.
This took months of corrections before it was useful, and it still gets things wrong every week. AI moved the judgment into a system that applies it every single time, instead of relying on whatever I remember at 11pm.
(This email started as an AI draft too. It had to get past everything above.)
If you’re using AI for your app’s listing copy, support replies, or onboarding emails, the pattern is the same. The model is the cheap part. The memory, the gates, and the feedback loop are what make it yours.
What have you handed off to AI so far, and where did it bite you? Hit reply, I’m collecting these.
Ohad