AI Writing Assistant

Professional editors do not fix a text in one sweep, and neither should you — or your AI assistant. Editing collapses when structural questions and comma questions compete for the same attention. The craft is to separate passes, each with one job. Pass one is structural: is the main point where a reader will find it, does the order of sections follow the reader's questions rather than your discovery process, is anything missing or duplicated? Polishing a paragraph that should be deleted is the most common way to waste an editing hour. Reverse-outline the draft — one line per paragraph stating what it does — and the structural problems become visible at a glance. Pass two is the paragraph pass: one idea per paragraph, the idea stated early, evidence following, transitions doing real logical work. Only pass three touches sentences: active voice, cut hedges, shorter verbs, concrete nouns. This maps directly onto how to use an assistant. Asking a tool to improve a draft runs all three passes at once, badly — it will polish sentences inside paragraphs that should not exist. Run separate requests: first "list the structural problems, do not rewrite anything", then paragraph-level feedback, and only at the end line edits.

Four edits account for most of the improvement in typical business and web prose, and they come straight from the working style guides — Strunk and White's "omit needless words", Orwell's "never use the passive where you can use the active", and the plain-language movement that turned both into checklists. Activate the verbs. Passive voice is not forbidden — it is fine when the actor is unknown or irrelevant — but default-passive prose hides who does what. "Mistakes were made" versus "we made mistakes" is a difference of accountability, not grammar. Related: hunt nominalizations, the verbs frozen into nouns. Make a decision, conduct an analysis, provide a recommendation thaw into decide, analyze, recommend, and each thaw saves words and restores motion. Cut the hedges. I think, perhaps, somewhat, it could be argued — one hedge is a nuance; three per paragraph is a writer asking not to be blamed. Keep the hedge only where the uncertainty is real information. Kill throat-clearing openers. "It is important to note that" adds seven words to any sentence and zero meaning. The same for "in today's fast-paced world" and every cousin. Then read one paragraph aloud. The ear catches rhythm problems, accidental rhymes, and triple abstractions that the eye forgives. This is also the fastest way to notice when an AI rewrite no longer sounds like you.

Before (37 words):
It is important to note that a decision was made
by the team to conduct a review of the onboarding
process, which is somewhat lengthy, in order to
potentially achieve an improvement in user experience.

After (15 words):
The team decided to review the onboarding process,
which is too long, to improve user experience.

Readers online do not read sentences; they sample paragraphs. Eye-tracking studies of web reading show an F-shaped scan: first lines get read, the rest gets skimmed. That makes the first sentence of each paragraph disproportionately powerful — and gives you a merciless test for structure. Read only the first sentence of every paragraph in your draft, in order. If that skeleton tells the whole story in the right order, the structure works; wherever it goes vague ("another thing to consider..."), a paragraph is hiding its point in the middle, where skimmers never look. The classic pattern is point-first: claim in sentence one, evidence and development after, and an ending that either sharpens the claim or hands off to the next paragraph. Academic writing tolerates build-up-then-reveal; the web does not. Two mechanical rules do surprising work. One idea per paragraph — the moment you write "also", check whether a new paragraph has begun without your permission. And make transitions carry logic, not lubrication: however, because of this, the exception is tell readers how ideas relate; additionally and moreover merely announce that more text exists. A draft whose paragraphs could be shuffled without harm has no argument yet — no amount of sentence polish will fix that.

AI assistants are excellent at the mechanical layer: grammar, agreement, punctuation, spotting the sentence you accidentally wrote twice, enforcing consistency of tense and terminology across a long document, and translating register ("make this formal") when the target is conventional. For non-native writers, this is a genuine equalizer — article errors and preposition slips vanish. The damage happens above that layer, and it is quiet. Voice flattening is the big one: every metaphor gets normalized, every short punchy sentence gets improved into a medium-length one, and after two full-text passes your writing sounds like everyone else's — fluent, warm, and anonymous. Meaning drift is subtler: a rewrite of "we will likely ship in March" as "we will ship in March" has changed a forecast into a promise, and rewrites of technical or legal text drift this way constantly. Worst is fabrication in expansion: ask an assistant to flesh out a thin paragraph and it will invent supporting specifics — plausible statistics, an example customer, a study — that you never provided. Every fact in an expanded draft that you did not put there yourself needs a source or a deletion. The protective habits: prefer targeted requests ("fix grammar only, change nothing else") over "improve"; ask for a list of suggested edits with reasons instead of a rewritten text, so you accept or reject each one; and diff the output against your original before using it. The diff takes one minute and catches both drift and inventions.

Riskier request:
"Improve this paragraph."
(rewrites everything, voice and meaning included)

Safer requests:
"Fix grammar and punctuation only. Change nothing else."

"List up to 8 suggested edits as 'original -> suggestion
 -- reason'. Do not rewrite the text. Do not add any
 facts, numbers, or examples that are not already in
 the text. Keep hedges like 'likely' unless they are
 grammatically wrong."

A sustainable division of labor: you own ideas, structure, claims, and voice; the assistant owns mechanics and acts as a first reader. In practice: draft fast and badly yourself, without the tool — a mediocre paragraph in your own voice is better raw material than a polished one in the model's. Then run the passes from this guide as separate requests: structural feedback first ("what is unclear, what order fails, what is missing — do not rewrite"), then paragraph checks, then grammar-only cleanup. Use the assistant as a reader simulator, one of the highest-value prompts there is: "What would a skeptical reader push back on? Where did you stop paying attention? Summarize my argument in two sentences." If the summary is wrong, the draft, not the reader, is at fault. Keep a hedge budget in mind when accepting suggestions, and re-read the final text once asking a single question: does this still sound like me, and is every fact one I can stand behind? Both failure modes — borrowed voice and borrowed facts — are ultimately attributed to the byline, and the byline is yours.