004: One File Makes Claude Write Like a Human

Human sounding AI content is not a prompting trick or a list of banned words. It comes from giving the model the rules once, in a file it reads before it writes a single sentence. This episode turns one LinkedIn post I saw from the couch into a small, repeatable system: I send Claude Code to research every tell that marks writing as AI, fold the findings into a single Markdown rulebook, and run that rulebook across everything I publish. It flagged almost 200 places to fix on my own blog. This article extracts the method. The session is the case study.

Why this matters

AI phrasing is everywhere now, and two audiences notice it. Real people feel it: the "it's not this, it's that" rhythm, the same ten words in every post, the tidy summary stapled to the end. The platforms feel it too. Google folded a scaled content abuse policy into core ranking, so low value content published at volume gets demoted no matter who or what wrote it. Content that sounds machine-made also quietly undercuts your product marketing, because the words meant to set you apart now sound like everyone else's. After fifteen years in product, the pattern I see with the founders I coach is that they reach for AI to publish more and end up sounding like the crowd that did the same. The fix is not writing less with AI. It is teaching the AI your bar once, so the output clears it every time.

#1: A list of AI phrases is rung one, not the fix

The honest version of how most people meet this problem: a list of AI tells shows up in the feed, you nod, you screenshot it, and by lunch it is gone. That was me last week, lying on the couch, looking at a LinkedIn post of phrases that instantly out an AI. The list was good. The list was also useless to me, because reading it changes nothing about what my AI produces tomorrow. The next time I generate a draft, the model has no memory of the post I liked.

That is the trap. A list is a one time hit of awareness. The work it is pointing at is a system that applies the awareness every single time, without me in the loop. The point was never the twelve phrases. The point is the layer that catches the twelve phrases, and the next twelve, after I have forgotten all of them.

The portable rule: when a good list crosses your feed, do not save it, build the thing that uses it. A rule you have to remember is a rule you will not apply.

#2: "Write like a human" is the instruction that does not work

The obvious move is to tell the model "write like a human" or "make this sound less like AI." It does not work, and it cannot, because the instruction has no content. You are handing the model an adjective and hoping it reverse engineers your taste. What comes back is the model's average idea of human, which is the exact center of the distribution that reads as AI in the first place.

The instruction has to carry the actual rules. Not "sound human" but "negation flips are banned, swap delve for look into, vary your sentence length on purpose, name a real date and a real number." Specific, checkable, ranked. The difference between the two approaches is the difference between a wish and a spec.

Approach	What you give the model	Result
Paste a list once	Awareness, for one chat	Forgotten by the next session
Tell it to "sound human"	An adjective	The model's average, which reads as AI
Hand it a ranked rulebook it reads first	Specific, checkable rules	Human voice as the default output

The portable rule: the model cannot act on an adjective. Give it a rulebook with severity, swaps, and examples, or expect the average back.

#3: Let the AI research its own tells, and keep the receipts

Rather than trust one post, I wrote a prompt that sent Claude Code to do deep web research on what actually reads as AI, what reads as human, and how Google and the detectors treat it. After a long pass across peer reviewed studies, detector vendor data, and editor guides, it came back with a document far longer than the post that started it: the sentence shapes, the cliché phrases, the overused words with plain swaps, and a section on the markers that make writing read as human. Receipts included, every claim linked to a source.

Keeping the sources matters because the tells move. "Delve" was the loudest signal in 2024 and started fading in 2025 once everyone learned it. A sourced doc can be re-run; a borrowed list goes stale and you never know when. This is also where I would do it differently next time: I would have run this research a year ago, before I wrote dozens of articles, instead of after. The platforms have been clear that the production method is not the problem on its own. As Google puts it:

Our focus is on the quality of content, rather than how content is produced.

The portable rule: have the AI build the evidence base and cite it, so you can refresh the rules instead of trusting a snapshot someone else took.

#4: Put the rules in a file the model reads before it writes

A rulebook sitting in a chat you closed is back to rung one. The step that makes it stick is turning the research into a single Markdown file and making the model read that file before it writes anything for me. Not a reminder I paste when I remember. A reference it loads first, every time, the way it would check a style guide.

So the research became the Human Voice Playbook: tells ranked by how strongly the evidence flagged them, every rule tagged HARD, SOFT, or JUDGMENT, a plain swap for each overused word, and a section on the human markers to add on purpose. Now the writing starts from the bar instead of drifting toward it. In my coaching experience, this is the same move that separates teams who keep their quality and teams who lose it as they scale: the standard lives in the system, not in one person's memory. A few hours of setup buys back the time I used to spend policing every draft by hand.

The portable rule: put the rules where the model reads them on its own. Voice you have to request is voice you will eventually forget to request.

#5: Audit what you already published, then stay honest

The new rulebook is also a flashlight for old work. I pointed it at my blog and asked it to list every hit without changing anything: file, line, the rule broken, the exact text, a suggested fix. It came back with almost 200 places to touch across articles I had written and was proud of. None of it was lazy writing. It was the tells leaking in unnoticed, one drifted sentence at a time.

Here is the part I want to keep honest, and it is the most important rule in the whole system. Every tell in that research also appears in real human writing. The rulebook is judgment, not a find and replace script. The pattern I see when people over-correct is writing that is technically clean and completely dead, every sentence sanded flat to dodge a detector. That is the wrong target. Do not write to beat a detector. Write so a real reader stays, and the detector takes care of itself. The almost 200 flags were a conversation with my own work, not a list of crimes.

The portable rule: audit the back catalogue, fix with judgment, and protect your voice over your score. A clean detector result on writing nobody finishes is a loss.

What is next

This system plugs into the build-in-public pipeline I set up in episode 001: one video, one article, one bundle, now with a voice bar wired through all of it. The three files I used are free in the download above. Take them, point them at one thing you wrote this week, and see your own number. If you want a second pair of eyes on how your content and positioning actually land with the people you are trying to reach, that is the kind of work I do in my product coaching. Subscribe so the next episode lands when it goes live.