I was eating breakfast and scrolling LinkedIn, which I tell myself counts as prospecting. A screenshot stopped me mid-bite: a prompt by Pia Szabo, 437,000 views, a long list of rules to make AI writing sound human. I write every day of this 365-day build with an AI in the loop, and I already run rules of my own, a banned word list plus a cliché blacklist that a critique pass reads before anything ships. I read her list twice and recognised my own posts in about half of it.
That stung a bit. I thought I had this covered.
So I tested it, on yesterday's post about the AI controller. Two drafts of the same article. Draft one under my existing rules. Draft two with hers layered on top. I read both and edited the one that sounded more like a person.
Draft one passed every check I had. No em-dashes, no banned words, scene first, UK spelling. It still read like a machine. When I counted it against her list I found eight uses of the "It is not X. It is Y." pattern in a single post. Eight. Three lists of exactly three items. Two sentences whose only job was to announce that the previous sentence mattered. I had no rule against any of it. My rules ban words, and you can't grep for a sentence shape.
I edited draft two and shipped it yesterday. Go compare if you like, the tells are easier to spot once someone hands you the list.
The whole stack, in case you want it
Her prompt is in the original post and you can paste it into any model as-is. The five rules from it that caught me, now permanent in my own setup:
- No corrective negation. "It is not X. It is Y." Say Y. Leave X out.
- No antithesis pairs. "Not that the humans disappear, but that the hours shift." Keep whichever half carries the meaning.
- No rule of three. The triplet is the machine's favourite rhythm. Two items, or four, or rewrite the sentence.
- No setup/payoff scaffolding. "Two things happened. The first was... The second was..." Fine to say two things happened. Then just tell them.
- No filler intensifiers. Actually, really, truly, genuinely. Cut the word and check the sentence. It survives.
Under that sits my older word layer, built from months of edits: zero em-dashes anywhere, and a flat ban on the abstraction nouns AI reaches for when it wants to sound strategic, the shape-lens-lever-unlock family. My test for a word is whether it would look at home on a consultant's slide. If yes, out.
The two rules I threw out
She bans sentence fragments. I use them all the time. Like this one. Take them out and the writing stops sounding like me, and sounding like me is the whole reason I do this by hand every morning instead of hitting generate.
She also bans hedging. My hedges stay, because they are usually the most honest sentence in the post. A while back I wrote that a CFO had never once opened an AI tool, then corrected myself in the next line: well, that may not be completely true, they probably used the free version of ChatGPT at some point. Readers trust the second sentence more than the first. Her prompt is tuned to produce a generic human. I need a specific one, with a Bermudian accent and a forestry degree.
The part I can't turn into a rule
Something I noticed during the test and don't have a tidy answer for. The lines people replied to in yesterday's post were the ones I typed myself, after the cull was finished. The damn-it-feels-good line when the month closed clean. The admission that a bug embarrassed me. No rule produced those. The closest I've got to a method: run the cull, then sit with the post for ten minutes and add whatever you'd say about it to a friend over coffee. Humans may not be as grammatically correct as AI, but to be sure, they sound more personal.
My blacklist file gained three entries this morning and is now 123 KB. It started in April as maybe ten banned phrases. Tomorrow is Day 141 and I already know the topic, which hasn't happened in weeks.
Day 140 of 365.