Getting rid of AI flavor isn’t something a set of prompts or an open-source skill can solve.

If you want to strip away AI flavor as cleanly as possible and make what you write sound more like yourself, you need to build yourself an iterative, compounding writing system. https://t.co/lSQahwtod2

Read This Article and You’ll Know Exactly How to Remove That Damn AI Flavor

Hey everyone, this is Jin Chenma.

There are already plenty of methods online for removing AI flavor from writing.

Combine a few “de-AI” prompts or open-source Skills, disable “first, second, finally,” ban notorious AI-isms like “not… but rather…”, cut back on parallel structures, add more colloquial language, break long sentences into shorter, more casual ones — that kind of thing.

These methods aren’t entirely useless. They can genuinely help remove some of the more obvious AI tells, making articles less stiff and a bit smoother to read.

But I later discovered a problem: some articles no longer show the classic AI tone, yet they still don’t feel like the author’s own writing.

That’s the crux of the issue.

Surface-level AI tells can be stripped away, but the author’s voice still doesn’t come through.

If removing AI flavor only stops at word choice, sentence structure, and formatting, you’ll likely just end up swapping one template for another.

Today, I want to share what I’ve tried when it comes to removing AI flavor, and what my ultimate approach looks like right now.

01 | My Understanding of “Removing AI Flavor” Has Changed Twice

My own thinking on this has kept evolving.

In the beginning, I also experimented with various prompting tricks to remove AI flavor.

Giving the AI a persona, specifying the level of colloquialism, restricting sentence patterns, telling it which words to avoid. These practices might work in the moment — they can turn an obviously AI-generated draft into a fairly normal-looking article.

But later I realized that adding these restrictions alone isn’t enough.

If the author doesn’t have their own opinions, the AI can only average out generic answers. If it doesn’t know how the author usually writes, it can only deliver a style that “most people would probably accept.”

So I changed my approach to: “author’s perspective + a detailed style guide.”

That worked a little better — the AI would produce content that seemed to resemble my own style.

But as I kept going, I found that the style guide was still just one layer.

A tutorial and an opinion piece, even written by the same person, differ in structure, pacing, and how material is organized. It’s hard to articulate which expressions an author dislikes through positive style descriptions alone. Even if you spell out every rule, the AI won’t necessarily execute them perfectly in the first draft. And if you don’t record why you made edits after revising an article, the same problems will repeat next time.

My current approach is to build a writing expert system.

It doesn’t just manage how to write — it also manages who the author is, what qualifies as a successful article, what must never appear, how to review the first draft, and how each round of revision becomes new rules for the next piece of writing.

02 | AI Tone Can Be Removed, But That Doesn’t Mean It Sounds Like You

I’m not against the “de-AI” prompts or open-source Skills available online.

For one-off writing tasks, prompts can tell the AI who the audience is, how long the piece should be, and what format to deliver. A well-made open-source Skill can also organize a generic writing workflow, checklist items, and delivery standards, saving people a lot of trial and error.

They’re well-suited for producing first drafts, but they can’t deeply solve the AI-flavor problem.

Because they still don’t know why you value a particular detail, which parts you’re willing to elaborate on, or how you’d respond to dissenting opinions. They also don’t know that there are certain things you refuse to write even if they’d attract more clicks.

An author’s true distinctiveness isn’t whether they say “honestly,” nor is it deliberately fragmenting sentences and stuffing in a few colloquial words.

It’s the author who decides what deserves to be expanded and what gets one sentence glossed over. Some things warrant definitive judgment; others must remain uncertain. As for experiences and identity — whether they can be written about and to what extent — there are clear boundaries.

These are things that other people’s prompts and open-source Skills can’t possibly know on their own.

03 | The Writing Expert System Solves Three Main Things

I now manage three things separately: foundational personal style, approaches for different article types, and content that must never appear. Together, they form a writing expert system. It mainly solves the three problems below.

First, who is the author of this article?

This doesn’t mean making the AI repeat the author’s bio in every article. It means letting the AI know what the author has been through, what perspective they habitually take, and what their relationship is with the reader.

More importantly, it needs to know the author’s factual boundaries. Which things can be told in first person, how to reference other people’s experiences — all of this needs to be specified in advance. The AI also can’t make judgments that go beyond what the author actually knows.

Without this, the AI can easily fabricate an experience for the author, add a stance, and then arrive at a conclusion that looks perfectly reasonable. The text might read smoothly, but the identity has already overstepped its bounds.

Second, how should this article be written?

This isn’t just about whether the tone matches — it also includes how a piece advances its arguments, when to give examples, where to pause, how long and short sentences work together, and where the ending should land.

Different article types need different standards. Tutorials need to clearly explain steps, results, and risks; opinion pieces need to take a stance and make it defensible. If you only define one generic personal writing style, you’ll end up turning every article into the same mold.

Third, what would this author absolutely never write?

It needs to prevent the AI from fabricating facts, and also prevent it from writing other people’s experiences as the author’s own. Those authoritative tones that don’t match the author’s identity shouldn’t appear either. Beyond that, it also covers expressions the author repeatedly uses and repeatedly avoids.

A lot of people building personal style only tell the AI what they like.

But what a person consistently refuses to write often says more about who they are.

Of course, you could stuff all three types of rules into one mega-prompt, but putting them in doesn’t mean the AI will execute them reliably every time, nor does it make them easy to maintain, reuse, or combine.

Prompts handle the article in front of you, but turning those prompts into a personal writing system preserves the author’s accumulated experience — so you don’t have to start from zero every time you open a new conversation.

04 | Multiple Reviews Are Required After the First Draft

Laying out all the rules clearly doesn’t mean the AI’s first draft is ready to publish.

Large language models still hallucinate, and they still miss requirements. When there are too many rules, the model might nail the structure but forget the tone; remember the personal style, then slip back into its own most familiar expressions in one paragraph.

So the first draft is just version-one material, not the final article.

If the first draft is a 60, the subsequent review rounds are what push it toward 80 or 90.

My current approach is this: after the first draft, I have Agents do several rounds of independent review. Each round handles one clear set of tasks, re-reads the corresponding rule files, and revises the same draft accordingly.

Round one checks whether the task is complete. Is the main thread clear? Did the structure go off track? Are the materials used correctly? Are any key arguments missing?

Round two checks whether this is something this author would have written. Did the personal style get executed? Does the article type match? Does the relationship with the reader align with the theme this piece is trying to express?

Round three specifically checks for AI tone and forbidden content. Which sentences just sound complete but carry no actual information? Which expressions are too polished? Is there any identity overstepping, fabricated experience, or writing patterns the author dislikes?

Each round can be assigned to a sub-Agent, which works better than having a single Agent review three times.

Every round must be strictly checked against the standards — don’t let the model that wrote the draft score itself a “no problems” based on its own impression.

Through these rounds of review, the article gets polished again and again. And in every round, which suggestions to accept and which sentences to keep is still the author’s decision.

Because review can surface problems, but it can’t make all the judgments on the author’s behalf.

05 | Writing Rules Must Continuously Iterate

A personal writing system isn’t a configuration you build once and never changes.

It should be continuously refined through every subsequent article.

Every review, every manual edit, generates new personal context. When revising, you should keep asking: what exactly was wrong with the original sentence? Why does it need to change? Is this problem a one-off occurrence, or has the AI repeated it many times already?

If the author always deletes a certain tone, add it to the banned expressions. If a certain article type keeps showing the same structural issue, adjust the corresponding type rules. If the author has formed a stable judgment on a category of things, save it — so the AI doesn’t give a completely opposite generic answer next time.

But this also doesn’t mean turning every local tweak into a permanent rule.

If rules only ever multiply and never get pruned or maintained, they’ll get longer and longer, conflict with each other, and the AI won’t know which rule to follow.

Only problems that recur, have significant impact, and can be reused next time are worth keeping. When rules conflict, merge them. When they’re outdated, replace them outright.

As this system is used over time, its core will evolve along with the author. It can record how things were written in the past, and also retain the reasons behind current changes — making every shift traceable.

In Closing

Removing AI flavor isn’t ultimately a simple prompt — it’s more like a loop.

First, you clearly define the author’s experiences, judgments, and boundaries, and let the AI write the first draft. Then you run several rounds of independent review, revising again and again. Finally, you feed the new problems exposed this time back into your personal rules.

The next article starts from a more accurate baseline.

This system of mine has been running for several months now. From my actual experience, my writing style has become more and more like me, and the quality keeps getting better.

If this article gets good engagement, I’ve decided to open-source my de-AI-flavor Skill.

Jin Chenma | Big tech programmer | 30 days to 10K followers, monetized over 10K | Continuously sharing AI money-making, programmer career pivots, and OPC insights | Contact info on profile: https://x.com/jinchenma_ai