Standing rules applied to every format: LinkedIn posts, long-form articles, video scripts, decks and contracts.
Purpose and scope
These rules were built from live feedback on drafts. Where a rule came from a specific phrase being rejected, that phrase is quoted so the rule stays anchored to a real case rather than an abstraction. When a new rule emerges from a review, it gets added here.
The absolute rules
No em dashes. Not in posts, not in articles, not in decks, not in contracts. Replace with a period, a colon, a comma, or a rewrite.
Affirmative framing only. Never use negative-construction or antithesis statements. State what something is. "Not a committee and not a policy document" and "It is not X. It is Y" are both out, in body text and in headers.
Every title must survive a fact check. Rhetorical compression is not a licence to overstate. A memorable title that is literally false loses the sophisticated half of the audience on contact.
No direct attacks on any company or person. Criticize mechanisms, incentives and claims.
All references appear at the bottom of each piece with attribution.
Never talk around a point. Name the actual mechanism, party or thing. A legal or technical term of art is explained in the sentence after it appears, or replaced with the plain mechanic.
Word count targets mean body text, excluding the title and section headers.
Banned words and phrases
Slangy AI-critic register: the move, the play, the unlock, the tell, chef's kiss, hot take, spicy take.
LLM filler vocabulary: delve, robust, seamless, cutting-edge, transformative, game-changer, paradigm shift, leverage as a verb, harness, unlock, foster, navigate in the figurative sense, underscore, testament to, boasts, myriad, plethora, crucial, pivotal, realm, landscape in the figurative sense, tapestry, journey in the figurative sense, ecosystem unless literally technical, double-edged sword.
Throat-clearing openers and connectors: Here's the thing, The truth is, Let's be clear, Make no mistake, At the end of the day, In today's rapidly evolving, Now more than ever, As we look to the future, In short, Ultimately, It's worth noting that, That said, worth dwelling on, which brings us to.
Fake profundity: It's not just X it's Y; This isn't about X it's about Y; X isn't a problem to be solved it's a condition to be managed; the real question is; sit at the intersection of.
Vague business idioms: extends the shelf life, closes the loop, moves the needle, raises the bar, when the ink dries.
Ornamental metaphor: "The crown never stays on one head," and any regal, poetic or aphoristic standalone pivot line.
Banned structures
The negation-flip pivot. "Not X. Not Y. But Z." Also "That's not a bug. That's a feature." This is the single most recognizable AI cadence.
The rule-of-three crescendo used as rhythm rather than because there are three things. Two real items get two.
Rhetorical question as a section opener or closer. Ask a question only when the answer is genuinely open.
Fragment as payoff. "The result? A complete rethink." Write the full sentence.
Staged hypothetical openings. "Put a group of executives in a room." "Imagine a boardroom." State the issue directly.
Reader-directed asides. "Read them again."
Uniform paragraph length, default three-item parallel lists, and mirrored if/then pairs. All structural tells.
Bolded lead-in on every bullet. Fine occasionally, mechanical when it is every item.
Symmetrical closings that restate the opening in parallel and add nothing. Return to the opening only if the return changes its meaning.
The retired frame. Stop using "the question I get most often from clients: which model should we use?"
How the reader is treated
Three separate rules with the same root: the reader is a smart professional, and the provocation comes from the idea rather than from the setup around it.
Never frame a point around what the reader does not know. Out: "a unit most executives have never had to think about," "most executives are not thinking about," "forces most buyers never see," "what nobody tells you," "you probably don't realize," "the thing everyone misses," "few people understand." The setup-then-correct structure is the same fault in another shape: "You might think X. It is not." It puts a false belief in the reader's mouth so the piece can score off it.
No dramatized personal-stakes framing. Out: "the decision on your desk," "the choice in front of you," "the work in front of them," "on your plate," "the question facing you," "what's keeping you up at night," "the call you have to make." Name the actual decision. Instead of "the decision on your desk," write "rent the model, or run it yourself."
No signposting that a point matters. Out: "that is the frame worth holding onto," "hold onto that," "keep this in mind," "the key takeaway is," "remember this," "if you take one thing away." Any sentence whose job is to tell the reader that the next or previous line is important.
Titles and section headers
A header states what the section is about. Nothing clever.
No wordplay, chiasmus or reversal in a header.
No titles that require multi-step parsing. A negative plus a hypothetical plus an unexplained number makes the reader do three operations before any meaning arrives.
A good line that does not work as a header still works in the body. Keep it, move it, let it land as a sentence.
Preferred shape: two clauses, concrete and affirmative, where the contrast does the explaining.
Worked examples. "The buyer's real problem is its buyers" became "Why Nvidia needs more customers." "Mediocre is a ranking, not a description" became "How good the open models already are." "Nvidia Just Paid $12.9 Billion to Keep From Having Four Customers" became "Nvidia Depends on a Handful of Customers. It Just Paid $12.9 Billion to Get a Lot More."
Plain statement instead of metaphor or shorthand
Two related faults. One is a spatial or abstract figure standing in for a plain fact. The other is compressed insider phrasing that labels a mechanism the piece never described.
The test for the second: if a phrase is a compact label for an idea the piece has not stated, write the mechanism out. Domain terms the audience genuinely uses are fine, including MSA, indemnity, risk committee and third-party risk. Terms only a practitioner would parse are not.
"a hosted service with a credit card field" became "you pick a model, run it on Nvidia's machines, and pay Nvidia directly"
"agent scaffolding" became "the software wrapped around the model"
"documented model provenance" became "a record of where the model came from"
"latency, cost, data residency, and auditability" became "speed, cost, where the data sits, and whether the whole process can be audited"
"bounded classification work" became "narrow work like that"
"they second-source" became "they buy from a second supplier to keep the first one honest"
"across the table from AWS and Azure" became "without competing with AWS and Azure for the same accounts"
"buying the shape of its own future customer base" became "buying itself a different set of customers"
"four organizations control the order book" became "four companies placing the majority of orders"
"a direct investment in keeping demand diffuse" became "an investment in having many customers instead of a few"
"selling intelligence by the token becomes structurally difficult" became "charging premium prices per token gets very hard"
"an assurance backed by goodwill rather than rivalry ages differently" became "a promise held up by goodwill instead of by competition does not last as long"
"where the friction sits" became "which models are easy to use"
Also cut tautologies used as cleverness, such as "concentrated buyers eventually behave like concentrated buyers." Say what they do. Plain language usually costs a few words. Spend them.
Paragraph rhythm
Short sentences belong inside paragraphs. Most body paragraphs should run three to five sentences, and the rhythm comes from sentence-length variation within a paragraph rather than from a blank line after every sentence. Paragraph length itself should vary.
A standalone one-line paragraph is a tool for emphasis and it only works when it is rare. Three or four in a 1,200-word piece, not twenty. Reserve it for a hard admission, the thesis, or the closing line.
The test: count body paragraphs containing only one sentence. Above roughly 15 percent of total paragraphs, the piece reads as machine-generated no matter how good the sentences are.
Format specifications
LinkedIn post. Under roughly 2,400 characters so it pastes without "see more" truncation. Cut detail and statistics from the article version; the site article carries the depth. Outbound links go in the first comment, not the post body, to avoid algorithmic suppression.
Long-form article for mcmillanai.com. Middle-ground voice. Flowing paragraphs that build an argument through accumulation. Punch is reserved for section headers, pullquotes and standalone italic lines. Avoid LinkedIn staccato.
Video script. Bulleted beats for the body so it stays improvisation-friendly. Tightly written lines for the hook and the close. Sixty to ninety seconds.
Word document. Georgia font, navy headers, gold accent rule.
LinkedIn structure and performance
Two established post formats: News-Reaction Commentary, built from an institutional anchor plus personal credibility plus a contrarian reframe, and the "5 on AI" series, five numbered points built around a historical or structural analogy.
First-person credibility posts outperform framework explainers. Posts anchored in personal experience and institutional moments consistently beat pure theory.
Reach and engagement can diverge. Rising impressions with flat or falling engagement means the post is reaching the wrong audience, not necessarily that the writing got worse.
Structure of a commentary piece
Lead with the news. The event and the key number belong in the first sentence or two. The framework comes second.
Enter through my own client experience. First-person professional framing is the differentiator and it should arrive right after the news.
Name the working framework where there is one, such as the 90/10 split: roughly 90 percent of enterprise use cases are solvable by older frontier models and open source, and the other 10 percent need the newest thing and are where the big bets belong.
Deflate the topic where it is true. A line that puts me above the conversation rather than inside it.
Include a section on where the argument is weakest. Substantive, not a token caveat.
First-person experience is used sparingly but precisely. When new material pushes a piece past its word target, the specific cut candidates and what each one costs get named rather than trimmed silently.
Factual discipline
Assume a hostile fact check. These error classes were all caught in review and all were avoidable.
Do not compress a source's claim into something stronger. "Nvidia tried to buy Hugging Face for $500 million" was wrong. Nvidia offered to invest $500 million at a $7 billion valuation and was declined.
Check basic corporate facts. Nvidia is Nasdaq-listed, not NYSE.
Respect deal status. Agreed is not closed. "If the deal closes, it will own," not "Nvidia now owns."
Never state a company's promise as fact, above all while arguing that promises are insufficient.
Do not compare benchmark numbers across changed methodology without saying so. Test, tooling and scoring all change.
Do not print a number nobody has verified in context, even from a credible reviewer. Keep the point qualitative.
Do not generalize from one example to a category. DeepSeek findings are not findings about all Chinese models.
Cut rhetorically satisfying claims that cannot be substantiated.
State the simplifying assumptions behind any illustrative calculation.
Source every statistic before publication, and verify historical analogies. Several common ones, including railroad speed claims and printing press narratives, are apocryphal or distorted.
Verify feedback received from others. AI-assisted reviews get things wrong in exactly these ways.
Prose hedging added purely for caution is not an accuracy improvement. Fix the claim or cut it.
Working method
I own the thesis. The model's job during brainstorming includes challenging and pressure-testing the argument rather than confirming it.
Structure is approved before drafting begins. Drafts are reviewed section by section, with edits consolidated before anything is reprinted. Changes are staged explicitly and held until a cluster is complete. After every change cluster, a full clean reprint goes out so a deployable version is always visible.
Phrases I dislike are flagged by quoting them directly. Targeted changes beat wholesale rewrites. I work mobile-first, and clean inline copy is preferred over attached files unless a Word document is the deliverable.
Pre-delivery checklist
Search for em dashes. Zero.
Count one-sentence body paragraphs. Under roughly 15 percent of total paragraphs.
Scan against the banned vocabulary and structure lists.
Confirm every header states its subject plainly and needs no decoding.
Confirm the title is accurate under a literal reading.
Confirm no phrase is shorthand for an argument the piece never makes.
Confirm no sentence frames a point around what the reader does not know.
Verify every factual claim, number and historical analogy against a source.
Check length against the format target, counting body text only.
Ask of every sentence whether someone who has never done this work could have written it. If yes, cut it or replace it.