TL;DR: Removing AI flavor breaks into two lines you can scan sentence by sentence: one for “does this sound like a machine,” one for “is this sentence actually finished.” Five items can be scripted and automated; the rest requires human judgment. But this method can only strip the flavor away. It cannot grow a human voice. That boundary is its honesty.

“The difference between the almost right word and the right word is really a large matter — it’s the difference between the lightning bug and the lightning.” — Mark Twain

It wasn’t until I’d revised my self-editing rules for the tenth or fifteenth time that I saw they were really only asking two questions. The first: does this sentence sound like a machine wrote it? The second: is this sentence actually finished?

The previous post covered how I turned de-AI-flavoring into a working process. This one covers that process itself: what it looks like, which parts you can hand off to a machine, and which parts you have to handle yourself. If you use AI-assisted writing, this is something you can take and adapt directly.

One framing note on how I think about writing. From an engineering and math perspective, language looks a lot like Lego: a finite set of words, a finite set of grammatical rules, and an infinite number of sentences you can build from them. That means writing has a side that can be broken down and improved piece by piece. Get fluent with how the blocks connect, and your expression becomes more precise one block at a time. What I’m laying out here is a map of how those blocks fit together.

The First Line: Does This Sentence Sound Like a Machine?

This line is about voice. AI writing has its habits. The most common:

Binary opposition: “It’s not X, it’s Y.” “Not just X, but also Y.” This is AI’s favorite sentence structure. The Washington Post analyzed 328,744 ChatGPT conversations and found that this negative-parallel construction appeared in 6% of exchanges in July 2025.

False subjects: “The data told the story.” “The market gave us the answer.” Inanimate things don’t act on their own. Replace them with a specific person or a specific action.

All-purpose fill-in-the-blank: “Master X and you master the core competency of Y.” Swap out the subject and the sentence still works. That’s a sign it’s saying nothing.

Performed vulnerability openings: “Honestly, I used to feel lost too.” A confession with no details is manufactured.

Ceremonial closings: “Let’s wait and see.” “Are you ready?” Slogans are not resonance.

Em-dashes: In current online writing, heavy em-dash use is almost a signature. It reads as AI.

From 2025 into 2026, new tells have surfaced: red-flag words (English: delve, tapestry; Chinese equivalents like 賦能, 打造, 值得注意的是), structural compulsion (every paragraph signposted with “here is the detailed analysis,” rigid “looking ahead” closings), and a particular false-sincerity register (the therapist tone, “that’s a perceptive observation,” or spending four hundred words on a two-sentence idea).

Word lists go out of fashion. Delve had its run, and then everyone consciously dropped it. So don’t memorize lists. Look at structure. Underneath all of this are two things: AI picks the most predictable word, so the prose flows without any friction or surprise; and sentence length stays consistent, so the rhythm goes flat. Track those two roots. That’s more useful than memorizing a hundred words.

The Second Line: Is This Sentence Finished?

The first line asks “does this sound like AI?” The second asks something completely different: has the sentence actually been written through to completion? Is it precise enough? This line came out of my own editing process. You won’t find it in most de-AI-flavoring guides.

I’ve organized it into ten kinds of moves, the ones I reach for most often when revising. Supply a missing subject. Complete the verb. Keep punctuation in its lane (commas connect clauses; the Chinese enumeration mark connects single terms only). Qualify absolute claims (“impossible to verify” becomes “difficult to verify reliably”). Make logical connectives explicit. Replace casual words with the precise written form. Use prescriptive language for rule statements. Standardize quotation marks. Cut adjectives that aren’t earning their place. Convert bloated noun phrases back into verbs (“make a decision” becomes “decide”).

Behind all ten, one underlying rule: every word has to be doing something, and every sentence has to be finished.

What I find interesting is that nearly all my own revisions land on this second line, not the first. The AI-flavor tells, once you’ve kicked them, don’t come back. But “is this sentence actually finished” is something you have to watch forever.

What Goes to the Machine, What Stays with You

Of these two lines, five items can be handled by a script with genuine accuracy: parallelism, em-dashes, comma clusters, short-sentence runs, and filler phrases. I’ve written these as a script that runs automatically before a piece goes out. Machines count reliably. Work like that shouldn’t depend on a second pass with human eyes.

The rest cannot be automated. False sincerity, false balance, whether a sentence is actually complete: these require reading for meaning. A script can do an initial scan and flag possibilities, but the final call is always human. So this is not a one-click solution. It’s a division of labor: the machine counts, the person decides.

Calibration Matters More Than the Checklist

A checklist gives you direction. The difference between using a checklist well and using it badly comes down to calibration.

The easiest mistake is with parallelism. It’s the top AI-flavor tell, but it can also be your signature move. The rule is one per piece: the instance that carries your central argument stays; the others get replaced. Don’t cut all of them. One parallelism is a blade. Seven is a verbal tic.

Some things can’t be cut mechanically either. The Chinese particle 了 is load-bearing grammar. Cutting it by the checklist will break the sentence. There is a floor to trimming: refining means removing words that aren’t doing work, not cutting to the bone. A metaphor, a loose aside: if it’s holding the sense of presence or the human register, keep it.

The Tools Only Do Half the Job

This needs to be said plainly: this entire method only gets you halfway.

Back to Lego. The blocks can fit together in very precise ways, but there’s a ceiling: they’re standard parts. No matter how clever the assembly, the result carries the look of Lego. And when everyone is working from the same set of blocks, build long enough and things start to look alike. Writing works the same way. When you’re working with the same tools, built on the same underlying model, the result, however clean, tends to converge toward something others have also written.

More to the point: engineering precision is not the same as presence. The blocks are shared. Anyone can use the same words and the same grammar. But which block you reach for, how you connect it, and why you connect it that way: that’s yours. Style lives in the choices, not in the pieces.

This method cannot give you a human voice. A clean piece of writing may be technically sound and still carry no warmth, no trace of you. What warms an argument is witness: a real experience that happened, which heats the idea from the inside, not more explanation layered over the top. The checklist can’t touch that. You have to write it yourself.

Why tools can’t reach that, I’ll take up in the next post.

Take This Method and Use It

I’ve organized the two lines, the ten sentence-completion moves, and the lint script into a general-purpose skill, stripped of my personal settings. I’ll release it so anyone can install it and adapt it to their own workflow.

Before you do, keep the last thing in mind: it only helps you shed what isn’t you. The part that is you, you have to grow yourself.