TL;DR — I divide revision into two tracks: first look for repetitive or empty expressions, then check whether each sentence communicates its meaning completely. A script can flag five types of formal issues, but I still make the final decision about meaning and style. This is a personal editing method, not a detector for whether an article was written by AI.

I had revised my self-editing rules well into the teens before I noticed that they were usually asking two questions: Has this sentence fallen into another familiar formula? And has it actually finished saying what it means?

The previous article explains how I turned removing AI flavor into a process. This article explains the process itself. If you also use AI-assisted writing, you can treat it as a starting point for revision and then adjust it to your own voice.

I sometimes think of language as Lego. Words and grammar can be taken apart and recombined, and greater familiarity with their connections can make expression more precise. But the analogy only goes so far. Writing also contains experience, context, readers, and judgment. It cannot be solved with one box of standardized pieces.

The First Track: Has This Sentence Fallen into a Familiar Formula?

This track begins with voice. The following are expressions that make me pause during revision. They are not forbidden constructions that must be deleted whenever they appear.

  • Repeated binary contrasts: “Not X, but Y” or “Not only X, but also Y.” Contrast can sharpen a point, but when every paragraph uses it, readers soon hear the same rhythm.
  • Vague subjects: “The data told the truth” or “The market gave us the answer.” Sometimes this is effective rhetoric. If it hides who made the judgment and on what basis, I put the person and action back into the sentence.
  • Universal fill-in-the-blank phrasing: “Master X and you master the core competency of Y.” If the subjects can be replaced without changing the sentence’s usefulness, I probably need to add something concrete.
  • Unsupported confession: “Honestly, I used to feel lost too.” The statement may be true, but it needs an actual experience behind it rather than serving only as a shortcut to intimacy.
  • Ceremonial endings: “Let us wait and see” or “Are you ready?” I ask whether this is a meaningful invitation or simply what appears when the article does not know how to end.
  • Overused punctuation and signposting: Runs of em dashes, every section announcing “the following is a detailed analysis,” or the same fashionable vocabulary appearing repeatedly can all be worth flagging.

Symbols, sentence patterns, and vocabulary cannot by themselves prove that an article was written by AI. Human writers also use formulas, while AI-assisted drafts may have gone through many rounds of careful revision. I want to improve the article in front of the reader, not attach a label to its author.

Word lists also age. Instead of memorizing a hundred red-flag terms, I more often ask whether the same syntax appears too many times and whether word choice and sentence length remain trapped in one rhythm. These are problems I have observed in drafts, not a complete account of how every model generates language.

The Second Track: Has the Sentence Fully Communicated Its Meaning?

After finding familiar formulas, I still need to inspect the sentence itself. A clean sentence may be missing its subject, action, or conditions. Making the prose shorter does not necessarily make the meaning clearer.

I often make ten kinds of adjustments:

  1. Add the necessary subject so the reader knows who is acting.
  2. Complete the action by stating what is being done and to whom or what.
  3. Clarify the hierarchy of punctuation instead of using enumeration punctuation to string complete clauses together.
  4. Narrow absolute language by replacing claims such as “cannot be verified” with the limitation the evidence can actually support.
  5. State the cause, condition, or transition clearly instead of making the reader guess at the logic.
  6. Replace vague conversational wording where precision matters.
  7. When explaining a rule, distinguish requirements, recommendations, and personal preferences.
  8. Check whether quoted language is the original wording, a paraphrase, or my own summary.
  9. Remove adjectives that add no information.
  10. Turn nouns back into verbs where appropriate, such as changing “make a decision” to “decide.”

Each adjustment depends on context. Sometimes “make a decision” needs to emphasize the process, and a short sentence is not automatically better than a long one. My basic measure is whether the reader can understand what the sentence says and how it connects to what comes before and after.

This second track takes much of my effort during revision. Avoiding a few habitual formulas is relatively easy. Continuing to state things completely is daily work.

Let the Script Flag Five Types of Problems, Then Decide What to Change

I once connected five categories of checks to a pre-publication script: contrast patterns, em dashes, runs of enumeration punctuation, runs of short sentences, and preset filler phrases. These rules helped gather passages that might deserve another look.

Counting and judgment need to remain separate. A program can count configured symbols and match sentence patterns, but what qualifies as parallelism, filler, or monotonous rhythm depends on the rules and sentence boundaries. A useful rhetorical choice may be flagged, while a difficult sentence may escape every flag.

A flag points to something that may need another look; it is not a command to revise automatically.

For questions involving meaning, sincerity, and style, I may also ask AI for suggestions. I still compare those suggestions with the intended meaning and the experience behind the passage. A tool can help locate a problem; it cannot establish that a story that never happened is true.

This article describes a method I gradually developed at the time. It does not report a newly retested detection accuracy, and there is no score that can guarantee that the AI flavor has been removed.

Proportion Matters More Than the Checklist

I often use “keep one central contrast in an article” as a personal revision default. It reminds me to reserve contrast for the point where it has the most force. It does not mean an article becomes wrong as soon as a second contrast appears.

Whether a sentence pattern stays depends on whether it serves a purpose in this article.

If several contrasts genuinely help the explanation, there is no need to remove them to meet a quota. Conversely, one contrast can still need revision if its content is empty, even when it fits the limit.

Some words cannot be removed mechanically either. In Chinese, the particle 了 may carry aspect, change, or tone. Frequency alone does not make it redundant. Metaphors, asides, and conversational phrasing deserve the same judgment. If they preserve a scene, rhythm, or necessary relationship for the reader, they have a reason to remain.

Refinement needs to preserve the actions, conditions, and tone that readers require.

Tools Can Organize the Draft, but Experience Still Needs a Source

Returning to the Lego metaphor, public words and grammar are available to everyone. The difference lies in which pieces I select, how I connect them, and why I make those choices. Cleaning up a sentence can help readers see those decisions, but it does not automatically produce a distinctive style.

For me, a sense of authorship often comes from real experience, concrete trade-offs, and judgments I am willing to stand behind. AI can organize material I provide and point out where details are missing. It cannot establish that newly invented details actually happened.

Real experience needs a source; it cannot be filled in with invented details.

That does not mean every article must become an autobiography. Sometimes a clear definition, an accurate source, or an honestly unresolved question is enough to give the article a position of its own. What I want to avoid is inserting unsourced times, scenes, or people merely to make the writing appear more human.

This boundary also leads into the next article’s discussion of authorial experience. Tools can participate in expression, but I still need to decide which statements can be published under my name.

Publish the Method First, Then Adjust It to Your Own Drafts

The original article said that a general-purpose Skill was attached at the end, then later said it would be released in the future. That could make readers believe a download was already available. The scope is clearer now: this article publishes the checking method. It does not include an installable package, and it does not present settings from another environment as a generally validated tool.

To put the method into practice, take one of your older drafts and work through the two tracks paragraph by paragraph. Record which flags genuinely help and which merely catch elements of your own habitual voice, then adjust the rules.

Rules are a starting point for revision. The final decision still has to return to what this article is trying to say.

Keep a checklist when it makes the meaning clearer. If it strips away the original proportion and judgment of the prose, revise the checklist too. That is the part I am willing to give to a machine, and it is also why I retain the final decision.

Four-panel comic: This is a personal editing method, not a detector for whether an article was written by AI. Symbols, sentence patterns, and vocabulary cannot by themselves prove that an article was written by AI. A flag points to something that may need another look; it is not a command to revise automatically. Whether a sentence pattern stays depends on whether it serves a purpose in this article. Real experience needs a source; it cannot be filled in with invented details. Rules are a starting point for revision. The final decision still has to return to what this article is trying to say.

November 2026 revision: This revision limits sentence-pattern and punctuation checks to editorial prompts, removes unverified precise statistics and the attributed quotation, and clarifies the boundaries among personal defaults, semantic judgment, and actual deliverables. The script was not retested, and the site’s writing rules were not changed.