These are my writing voice preferences
I taught the machine my voice, not the other way around
Read enough published writing this year and you start noticing that a lot of it sounds like the same person. The same dashes doing rhythm work, the same over-reached verb everybody noticed last spring, the same tidy colon in every title, the same bolded stub leading every bullet. That isn't a style anyone chose. That's a few hundred million people accepting a model's defaults and slowly losing their own ear.
The direction of that trade is what bothers me. The human adapts to the machine, so the machine's habits win by attrition, and the writing converges on one voice that belongs to nobody.
I run it the other way. I keep a written contract of how I write, and the machine has to meet it. Right now it's on version fourteen. Every rule in it started as a correction I made out loud while reading a draft that felt wrong, and it stays in the file with the date I said it. "No double dashes, ever" is rule one because I said that on June 8th and meant it.
Think of a jig clamped to a workbench. You cut it once, by hand, to your own measurement, and after that the saw can't wander. The jig doesn't build the furniture and it has no taste of its own. It just makes your cut repeatable when your hand is tired. The contract is my jig, and a linter enforces it, so a draft that breaks my own rules fails before I ever read it.
The strongest objection to any of this isn't that a machine touched the words. It's three sharper things. The training data behind these models was taken from writers who never agreed to it, and writing carefully doesn't undo that. Real editors and writers are losing work, and my being a one-person shop that never had editorial staff doesn't make that sting less for them. And the flood of machine-shaped prose is degrading the commons for everyone who reads.
I can only answer the third one. Everything here exists to fight homogenization, and a voice contract that tells the model no, fourteen versions deep, is the opposite of accepting its defaults.
The machine drafts sentences. Not just polish, not just a proofread at the end. It researches, it proposes, and it argues with me. Then I direct it, overrule it, and own what ships. On one working day it tried to put "you" into a title where my own rules forbid second person, tried to ban a highlight I actually want, and tried to call italics a weak device when italics is exactly how I cue myself to stress a word aloud. I overruled all three, and all three corrections are in the contract now, where the next draft has to walk past them.
An editor imposes their own ear on your prose, which is what editors are for and why they're worth hiring. My linter imposes nothing but what I already said. The words that come out are more mine, not less.
You don't need my rules. You need yours, written down somewhere a tool can check them. Start with one, the tell you hate most in your own drafts. Write it as a plain sentence, put it where a machine can read it, and refuse the next draft that breaks it. That's the whole method, and the jig only ever has to fit your own hand.
The method
Three ideas do most of the work, and each one came from noticing a draft that felt wrong before knowing why.
The read-aloud test
Write what you would say to a listener, not what looks right on paper. Nobody hears an em dash, a colon subtitle, an arrow, or a clipped fragment, which is why those read as machine made. The test also has a practical payoff: an article that survives being read aloud can be recorded without a rewrite.
Register decides the rules
An essay and a reference catalog are different jobs. Essays carry almost no formatting and let sentence construction do the work. Reference pages may use the full scanning toolkit, since a reader arrives at those to look something up rather than to be persuaded. Applying essay rules to a catalog damages the catalog.
Formatting serves three readers
The content has to stand alone with every mark stripped, because quality writing carries its own emphasis. Bold exists for the scanner who never committed to reading, and it has to form a skeleton worth scanning. Italics exist as the cue for reading the piece aloud later.
The rules
Rule set version fourteen, last changed July 25, 2026. The eleven prose rules are judgment, applied while writing and cited by id during review. The word lists below are mechanical, and a linter fails a draft on them before anyone reads it.
The prose rules
- No em dashes or double dashes. Ever.
- No AI cadence. No fragments as style, no arrow chains, no bullet essays.
- First-person ownership, granted per surface and never inserted to satisfy a rule.
- Aphorisms stay verbatim and punchy. Describing an aphorism flattens it.
- Banned vocabulary, listed below.
- Verdicts called flat, without hedging.
- Concrete numbers over adjectives. A missing number is usually a missing claim.
- Narrative shape is the engineering story: wrong theory, the evidence that broke it, the correction, the lesson.
- Warm, not clinical. Contractions, and a willingness to admit frustration.
- Substance under every slogan. A quotable line owes the next sentence.
- The technique is the brand. Names stay stable across repos and articles.
Banned outright
A draft containing any of these fails. Most arrived as a single correction against a single sentence, and the hype absolutes came in together on July 25 because an empty superlative asserts volume where a fact belongs.
journey, journeys, interestingly, delve, delves, delving, delved, thrilled, excited to announce, excited to share, I'm excited to, we're excited to, reshaped my approach, in today's landscape, no one saw coming, changed everything, like never before, never be the same, against all odds, the rest is history, single-handedly, little did they know, what happened next
Flagged for review
These raise a warning rather than a failure. Some are AI-cadence tells, and the last few are formal constructions where the contraction sounds more like me.
seamless, seamlessly, supercharge, revolutionize, tapestry, moreover, furthermore, cutting-edge, let's dive, dive into, deep dive, game changer, game-changer, in the world of, in the realm of, it's worth noting, needless to say, at the end of the day, buckle up, stay tuned, without further ado, in conclusion, paradigm shift, unlock the power, unlock the full, elevate your, harness the, cannot, should not, nothing short of, forever changed
Emphasis
- Bold is structure. It marks what matters for the scanner, and the bolded text read alone has to deliver the core message.
- Italic is prosody. It marks the word to lean on when reading aloud, and a term mentioned rather than used.
- Both marks on one word is allowed, and rare.
- Highlight is the second scan tier for a reader with ten more seconds, and it never carries meaning by itself.
- No uniform stub grids. A bold label repeated down a list is structure wearing the wrong device, and it belongs in a heading.
- Capitals for acronyms and literal constants only. No underline except links.
- Titles and excerpts carry no markup, since they travel into feeds and machine-readable contracts.
- Register sets the budget. Essays carry almost none of this; reference pages may use the full toolkit.
Also enforced
- Never announce the virtue you are about to perform. State the thing.
- Write what you would say to a listener. Nobody hears a colon subtitle.
- Ground narrated events in a real actor. No disembodied abstract noun as the doer.
- No manufactured staccato. Short sentences are reserved for verdicts and aphorisms.
- End on an executable prescription, then one closing image.
- Internal codenames and client financials never appear in public.
Receipts
- Rule set at version fourteen, first entry June 8, 2026. Each entry carries a date and the correction that produced it.
- A deterministic linter runs on every draft and reports violations by line, with code blocks, tables, and frontmatter masked so examples never false positive.
- Long-form essays also pass a cross-model verification panel before publish. The panel result, including per-verifier scores, ships in the article itself.
- The corpus was swept against the contract on July 25, 2026, which removed 127 em dashes across 28 articles and rewrote 20 titles to read aloud.
Disclosure
StoneyTECH.net is independent learning and portfolio work. It is not affiliated with, endorsed by, or representative of any employer. The public material here does not use employer confidential information, employer systems, unpublished work product, proprietary customer data, or private original work. It makes no claim to original research; anything genuinely original stays out of the public corpus.
Preferences, not a standard. Take the method and leave the rules.
It's my voice.