I use AI to write. And I'm not ashamed of that.
The ideas are mine. The structure is mine. The experience that makes a technical article worth reading is mine. Kiro AI helps me tighten and compress it. That's co-authoring.
Here's the problem. Let AI write in its default voice and your readers will clock it. "In today's rapidly evolving landscape." "It is crucial to note." "Significantly enhance." The long dashes — everywhere — for no reason. "It's not X — it's Y." "Not this, that." Readers who know can't unsee it. They'll skip yours before they start.
Your ideas can be original and a reader trained on a year of AI slop will still skip it. They see the pattern and assume nothing's underneath.
The people who know how to fix this weren't AI researchers. They were copywriters.
I'm a cloud architect, not a copywriter. Before AI, I worked with human editors: weeks of drafts and rewrites, and the articles came out better every time. When AI became my writing partner, speed went up and quality went down. The editors' rules were in my head, not in the agent's instructions. So I put them there.
Stop sounding like AI, then learn to hold a reader
First I copied the Wikipedia "Signs of AI writing" lists into my instructions. The drafts stopped saying "delve," "robust," "seamless." They were still boring. Passing the detector isn't the goal. Holding a reader is.
So I went to five copywriters (Ogilvy, Sugarman, Zinsser, Halbert, Deutsch) for five rules that stack, each fixing what the last one exposes.
- The slippery slide (Sugarman): every sentence exists to make you read the next. Keep first sentences short; open a gap the reader has to close.
- Kill the throat-clearing (Zinsser): the first sentence of a paragraph often just warms up the writer. Delete it. "It is worth noting that deployment times improved significantly" becomes "Deployment times dropped from 20 minutes to 3."
- Show, don't explain (Deutsch): put a picture in the reader's head instead of a summary.
- The Ogilvy test: read it aloud; if you wouldn't say it to a colleague at a whiteboard, rewrite it. Nobody says "organizations leverage cutting-edge solutions." They say "we switched to X and it halved our deploy time."
- Loss framing beats gain framing: name what the reader loses by doing nothing, and lead with it.
Show, don't explain is the one that changed the output most.
Before:
"The integration was unreliable and caused frequent production incidents."
After:
"Last Tuesday the upstream team changed their payload schema without telling anyone. Forty minutes of downtime. The on-call got paged at 2 a.m. for something a single synthetic event would have caught."
The first reports. The second puts you in the room.
Why AI drifts toward mediocre
This isn't only my impression. RLHF, the training step that rewards pleasant, agreeable answers, narrows a model toward one safe register. Kirk et al. measured it (ICLR 2024): RLHF-trained models produce less diverse output than the same models before that step.
That's the current every draft drifts back into. The craft rules are the counterweight. My last six LinkedIn posts used them; one hit 22,000 impressions, and the comments were about the technical claims, not the writing.
It's not a one-shot fix
The skill doesn't produce perfect output on the first run. I still edit: cut a paragraph that explains too much, rewrite an opening that starts with context instead of the point. I used to do that for ten rounds. Now it's two or three. It raises the floor; it doesn't replace the editor. The agent is a co-author who needs direction, not a finished-content machine.
Sources:
- Kirk et al., "Understanding the Effects of RLHF on LLM Generalisation and Diversity" (ICLR 2024)
- Wikipedia, Signs of AI writing
- David Deutsch, interview on copywriting in the AI era
- Sugarman, Zinsser, Ogilvy, Halbert, Hemingway on craft.
Try it
One command:
npx skills add vidanov/writing-craft-skill
Works with Claude Code, Cursor, Kiro CLI, Codex, and 50+ agents. For ChatGPT or Claude Projects, copy chatgpt/PROMPT.md into your custom instructions. No CLI needed.
AI doesn't make writing generic. Generic writing makes AI generic. Teach it the craft instead.
https://github.com/vidanov/writing-craft-skill
MIT. Do whatever you want with it. If it saved you an editing round, a star ⭐ helps the next person find it.
Top comments (7)
just quick one have you verify the output of this experiment with ai detector?
After your request I tried it with zerogpt.com for the current article, with 14%
But the whole story is not about pass the detectors, it is more about how to create better content.
And this is GPTZero 😀
its superb hats off to your work sir.......!!!!
That's a good question.
I'm going to give this a whirl. I personally have had significantly better results by teaching AI through patterns, rather than explaining. Funnily enough, that's a major point of this. Why would AI (a pattern matching/producing tool) work any differently?
I have dramatically improved AI's ability to write code, by showing it how to write code, rather than using language to "create rules".
I suspect that this will help a good bit, but is likely to find its own ruts of common approaches that may be undesirable.
Thanks.
Thank you, Mike! Let me know what you notice after a few runs.