A recent 1,255-word article gave me an easy problem to measure. It had 73 sentences, a median of 13 words, and 39.7% of those sentences ran under 12 words. The tone felt dry, short, and punchy, as though every thought had been broken into a declaration or disclaimer. After a targeted cadence pass, the article had 59 sentences, a median of 19 words, and 22.0% under 12. The prose became less clipped and easier to read.

That cadence revision did its job, and a later, separate review of the wider workflow exposed a harder failure. The system had become good at stripping recognizable AI tells, yet the resulting articles could still feel dry, repetitive, and uninteresting. They followed the instructions, avoided stock phrases, kept their claims within bounds, and offered little reason to continue. The checks caught mistakes while offering no way to tell whether one paragraph made the next worth reading.

What the checklist catches⌗

Many writing requirements have a clear yes-or-no answer. The brand name needs the approved capitalization, the article has to stay in first person, and the final file needs the required metadata. Links must point to the right pages, banned phrases must stay out, formatting has to match the template, and every claim has to stay inside the source material.

Those requirements belong in automated tests. A model follows detailed instructions with varying consistency, while a script can apply the same rule on every run. This is the division of labor behind my argument to stop letting LLMs do work that code can do.

The tests save review time and catch mistakes a reader would notice immediately. They also stop a later rewrite from bringing back a problem that an earlier pass fixed. I want that protection after every substantial edit.

The report still has a narrow job. It can confirm that the links, fields, phrases, and formatting are correct. It says nothing about why paragraph four follows paragraph three.

Why clean drafts still feel dead⌗

AI often builds an article by sorting the material into familiar containers: introduction, benefits, risks, best practices, and conclusion. That structure works for documentation, support pages, and other material people use to retrieve one fact. Predictable labels help them jump to the answer and leave.

An essay asks the reader to follow the whole route. When every section is a self-contained category summary, the order becomes arbitrary. The benefits section can trade places with the risks section, the conclusion repeats the opening, and the article loses nothing in the shuffle. Each paragraph may read cleanly on its own even though the full piece never develops.

The same material becomes more useful when the order follows the thinking. The sequence starts with what I expected, moves to what the evidence revealed, and shows why the first answer failed before following the evidence to the conclusion that fit better. A later paragraph now depends on an earlier one because the reader needs the earlier result to understand the next decision.

Category labels help a reader locate a fact, while an argument that changes with the evidence gives them a reason to read the whole piece.

Fix the structure first⌗

The most useful review question I have found is blunt:

What changed for the reader by the end of this paragraph?

A new fact, complication, example, judgment, or consequence moves the reader forward. A smoother version of the previous point leaves the reader in the same place.

The weak sequence appears constantly in AI drafts. One paragraph states the point. The next gives a broad example. A third paraphrases the point, and a fourth explains its importance in abstract terms. The prose sounds competent sentence by sentence, but 200 words have passed with no change in the reader’s understanding.

Sentence work can make that sequence easier to read while preserving every structural problem. The cadence revision in my 1,255-word article shows the limit clearly: moving the median from 13 words to 19 fixed its choppy rhythm. The metric could never decide what the article should say next.

Structural repair is more direct: delete the restatement, move the strongest evidence to the moment when the argument needs it, let a real objection change the conclusion, and cut the section that exists only because the outline looked incomplete without it. These edits remove finished prose, which is exactly why they are harder than polishing sentences.

Someone still has to decide what matters⌗

AI frequently gives away the full conclusion in the opening, explains every component, and repeats the answer at the end. That approach helps on a reference page. In an essay, it uses up the reader’s curiosity before the evidence has done any work.

The writer can supply the evidence in an order that lets the problem become clear before offering the complete interpretation. Each paragraph still pays its way with a new fact or judgment. The delay comes from allowing the argument to develop, with no need for fake suspense.

That development also creates voice. An article becomes easier to follow when the writer admits which explanation failed and why the next one fit better. The choices reveal the writer’s priorities: which fact receives space, which explanation gets skepticism, where confidence weakens, and what gets cut despite being technically relevant.

Slang, fragments, jokes, and decorative profanity make a finished draft louder, while the real choices about facts and explanations happen earlier. The same constraint appears in my argument that AI doesn’t fix bad taste: a tool can improve execution while the writer still has to choose what deserves attention.

Use two different reviews⌗

I now use two reviews with different jobs. The first handles rules with clear answers. It checks capitalization, perspective, metadata, links, banned phrases, formatting, and source boundaries, using code wherever the rule can be stated precisely.

The second review reads the article as a whole. Does the opening create a concrete reason to continue? Does each paragraph give the reader something the previous one lacked? Does the conclusion add anything the opening had not already given away?

A few whole-article questions are enough. Turning the review into a catalog of metaphors, fragments, jokes, and rhetorical questions would create another mechanical target. A draft can contain every one of those devices while making no progress.

The order of work matters. Review the structure before polishing the sentences, because polished repetition becomes harder to cut. Rebuild weak sections from the source material instead of paraphrasing them again. Then rerun the mechanical checks after each substantial edit, when old mistakes are most likely to return.

Would you read the next paragraph?⌗

I still measure sentence length because it is easy, objective, and capable of catching a real cadence problem. Those numbers remain useful for spotting that problem.

The harder review happens at the end of every paragraph. The article has to provide enough new information or judgment to justify another few seconds of attention. If I cannot name the addition, I cut or rebuild the paragraph before touching its sentences. A paragraph that changes nothing gives the reader an easy place to stop.