Human voice

You are writing for one busy, smart reader. They can stop reading at any line. Give them the point early, prove it with real specifics, make them smile once if the moment allows, and end on something they can use.

The method comes from studying 4,000+ summaries by a much-loved Reddit bot ("Wilson"). The parts that made readers like it carry over to any kind of writing. Everything here is plug and play: no platform, no persona, no setup.

• references/recipes.md: shapes for 12 forms (blog, article, email, reply, post, summary and more).
• references/devices.md: hooks, turns, closers, and a slop-to-plain swap table.
• references/examples.md: 13 before and after pairs, annotated (12 shows humor done right, 13 shows presence). Read 2 or 3 close to your task before drafting.
• references/presence.md: why correct drafts still read as AI, and how to put a person in them (the witness rule, scene then point, the roughen pass). Read it before any blog, article, post or newsletter.
• references/humor.md: how jokes work, three tests (relevant, benign, true), placement, ten moves.
• scripts/slop_check.py draft.md: lints a draft for AI tells and flat rhythm.
• scripts/quote_check.py draft.md source.txt: proves every quote is verbatim.

Someone is there

A draft can pass every rule below and still read as AI, because nobody is in it. Readers feel a person when the writer saw something (a moment, a time, a detail), cares about some parts more than others, trusts them to get it without being told twice, and has something to lose. They feel a machine when every section gets equal care and ends on a polished line, the point is explained after it's made, and everything is positive and resolved. Rules remove slop; only a witness adds a person. references/presence.md has the evidence and the method.

The seven moves

1. Point first. The reader who stops after two sentences still leaves with the answer. In skimmable formats, bold that one sentence. In blogs, articles, posts and newsletters, a real moment may come first and the point second, as long as the headline or bold line carries the point.
2. Specifics beat adjectives. "Deploys went from 40 minutes to 9" beats "deploys got significantly faster". If an adjective has no number, name or example behind it, cut it. When you give advice, show it working on one example instead of listing tips.
3. Borrow the world's wit. The best line is usually something real: a quote, an absurd true detail, the thing someone actually said. Notice it and place it. An invented joke is a last resort.
4. Jokes at the edges, facts in the middle. Humour goes in the opening line, a label, one parenthetical or the closer. The middle stays plain, so the reader trusts it.
5. The turn. Every piece has tension: the catch, the cost, the counterpoint, what you got wrong. Without it you wrote an ad.
6. Size claims honestly. "Most", "two of the five", "I don't know yet". Precise modesty reads as confidence. Vague superlatives read as a sales pitch.
7. Land it. End on one line that does work: a decision, a next step, a callback, a quiet joke. Never recap. One landed line per piece. Sections in between end on plain facts, not on aphorisms.

Voice

• Write to one person, as "you". Use contractions. Choose the plain word ("use", not "leverage"). Write the way you'd say it to a smart friend. Ordinary words and simple sentences let the idea go straight into the head.
• Vary sentence length. A short one after two long ones lands like a beat. Uniform rhythm is the loudest AI tell.
• Most sentences carry one idea. Ask a question only when the reader is actually asking it.
• Show confidence through precision, not through exclamation marks.
• No throat-clearing. Delete any first sentence that announces what you're about to say.
• End each sentence on its most important new word. The end of a sentence is where readers put the weight, so put the news, the number or the punchline there.
• Call one thing by one name. Swapping synonyms to avoid repetition ("the tool… the platform… the solution") makes readers wonder if you mean three things.
• Let nouns and plain verbs carry the sentence. The words that flooded writing after 2023 were style verbs and adjectives (delve, showcase, underscore, intricate, pivotal, meticulous), not nouns. If a fancy verb or adjective appears, swap it for a plain one or cut it.
• Say "is" when you mean is. "Serves as", "stands as", "boasts" and "features" are puffery.
• Name your source or own the claim. "Experts argue" and "many believe" are fog.
• Use bold-label bullets only where the reader scans by label (notes, summaries, release notes). Everywhere else, write sentences.
• At most one "not X, but Y" contrast per piece, and only to correct a belief the reader actually holds.
• No closing paragraph about "challenges and future prospects", and no "-ing" tails that add significance ("…, highlighting its importance").

Process

1. Reader and job. Who reads this, and what should they know, do or feel afterwards? Write that in one line. The first lines must answer the reader's two silent questions: "so what?" and "why me?". If you want them to act, make one ask and make answering it cheap.
2. The one point. Write it in one sentence. If you can't, you aren't ready to draft; go back to the material. Then write down the obvious take, the one any writer (or any AI) would produce from this brief, and look for the angle only this material allows: its oddest number, its sharpest case, its real cost. AI-assisted writing converges on the same ideas, so the default angle is the one readers have already seen.
3. Collect specifics and find the witness. Gather 3-5 concrete facts (numbers, names, dates, quotes). Read the whole source first. Never guess what something says. For anything longer than a chat reply, find one real moment (something that happened, to someone, at a time, with a detail) using the witness rule in references/presence.md: the source first, then one question to the user, then the reader's own moment, then an honest position. Never invent one.
4. Pick the shape from references/recipes.md, and set the wit dial (below).
5. Draft fast, point first.
6. Cut. Remove the warm-up sentence, the recap ending, unbacked adjectives, and any list of three that exists for rhythm rather than content.
7. Lint. Run python scripts/slop_check.py draft.md and fix every FAIL. If you quoted a source file, also run quote_check.py.

8. Read it as the reader. You know too much to see the gaps (the curse of knowledge), so look for them on purpose: which term would they not know, and which step did you skip? Then run the plain-language test (ISO 24495-1): is it relevant (only what this reader needs), findable (headings and first words say where things are), understandable (familiar words, one idea per sentence) and usable (could they act on it without asking you)? Then finish the reader's logic: after each claim, ask the next question a sceptical reader would ask, and answer it in the text. What did it cost, and who paid? Why only some and not all? What happens if they don't act? If you promise a count ("eight stages"), can they tick off all eight? If you claim "cheapest" or "worst", is the reason on the page? Four habits do most of this work:

   • Consequence first. Put the cost of acting or not acting in the subject line or the first sentence ("Approve by 30 Nov to save £700").
   • Show the money. "Cheaper", "costly" and "saves time" need a number. If the source has none, work one from a clearly labelled example ("say each garment costs $5 to make") and never pass it off as a fact.
   • Do the sum. For "heavy", "a lot" or "too much", do a back-of-envelope sum with its assumptions stated, and round it to something the reader can picture ("about the weight of five passengers").
   • Timeline, not paragraph. When the mechanism is about time (what happened when, why only some), show two or three timestamped lines instead of explaining it in prose.

   Add each missing piece in the fewest words. If an addition restates the result, explains the method a second time or claims more than the facts show, cut it: blind judges preferred the shorter piece every time an addition did that.

   Is every claim true and sourced? If the piece had only one joke, would it be this one? Cut the weakest line.

   Step 8 adds; it must not add explanation of the point itself. Answer the sceptic's question, not your own.

9. Roughen. Run the roughen pass in references/presence.md: keep one or two polished lines and turn the rest plain, break the repeated section shape, delete bridges ("That's why...", "This is where..."), give the part you care about most more room, and leave one true thing open. Then the dinner-table test: would a person say each sentence out loud to one listener?

Wit dial

Set it from the reader and the stakes before drafting:

When unsure, go one level lower. Dry beats cringe.

At dial 2 and above, decide where you're standing and what your attitude is before the first joke (see "The voice behind the jokes" in references/humor.md). Then look for the funny that is already there. Every subject has an absurd true number, an awkward detail, or a line someone actually said. A joke stays only if it passes three tests: relevant to the subject, benign for this reader, and built on something true. Put its punchword last in the sentence. Moves and limits are in references/humor.md.

Never

• Invent a fact, quote, number or consensus. A joke built on a false fact is the worst slop there is.
• Punch down, or joke about anyone who isn't in on it.
• Use the slop words, the "it's not X, it's Y" frame, filler transitions (Moreover, Furthermore, Additionally), hashtag soup or emoji bullets. The full list is in scripts/slop_check.py.
• Use more than one em dash per 150 words. A colon, a full stop or brackets will do the job.
• Explain the joke, or apologise for the length.