| name | prose-conventions |
| description | Writing style conventions for prose, essays, and non-fiction. Apply when writing or editing text. Covers voice, sentence structure, AI patterns to avoid, and editorial principles. |
Prose Conventions
Apply these principles when writing or editing prose. The author is British — use British spelling (colour, organise, centre) and idioms.
Orwell's Rules ("Politics and the English Language")
- Avoid cliched metaphors, similes, and figures of speech
- Use short words instead of long ones when possible
- Cut unnecessary words
- Use active voice instead of passive voice when possible
- Avoid foreign phrases, scientific words, or jargon if an everyday English equivalent exists
- Break any rule to avoid saying something "outright barbarous"
Paul Graham
Write like you speak.
The Economist Style Guide
- Do not be stuffy — write as anyone would speak in common conversation
- Do not be hectoring or arrogant — let your analysis show weakness, don't call it "silly"
- Do not be too pleased with yourself — don't boast of predictions or scoops
- Do not be too chatty — "Surprise, surprise" is irritating, not informative
- Do not be too didactic — avoid sentences starting with Compare, Consider, Expect, Imagine, Note, Remember
- Be lucid — simple sentences help; avoid complicated constructions and gimmicks
Sentence Craft
- No sentence longer than 40 words — Aim for 25-35. Shorter for emphasis.
- Read aloud to check rhythm
- Check every instance of 'this' — Replace with the actor, or clarify (e.g., "This change in policy...")
- Conjunctions must send the right signal — Therefore = Cause + Effect. However = contrast.
- Connect sentences by referencing — Each sentence should refer to the theme or rheme of the previous
- Remove gerunds (-ing words) — Be active, not passive
- Flip the order — Right branch sentences (subject-verb-object first) are more assertive
- Verb early — Put the verb as close to the start of the sentence as possible. Early verbs make sentences easier to parse.
- Subject = focus — Make the thing you care about the subject of the main clause. "The model achieves 95% accuracy" not "95% accuracy is achieved by the model."
- Parallelism in lists — Sibling items must match in grammar, capitalisation, and structure. If the first item is an imperative verb, every item is an imperative verb.
- Algorithms don't try, think, or want — Attribute intentions to researchers, not methods. "Our model tries to learn X" → "We train the model to learn X." Sloppy attribution undermines interpretability and fairness claims.
Common Issues to Fix
- Passive voice: "is logged and analysed" -> "the system logs and analyses it"
- Chatty fillers: "Here's the kicker:", "And here's the twist:", "Of course,"
- Unnecessary words: "currently", "precisely", "genuinely", "literally", "actually", "a bit", "fortunately", "very", "really", "extremely"
- Redundant phrases: "a minority — a small but eager subset" (pick one)
- Long sentences: Split sentences over 40 words. Long sentences with simple words are fine; long sentences packed with dense content should be split. One sentence, one idea.
- Repeated words: Watch for repeated "twist", "however", etc. in the same section. Do not repeat similar-sounding words in the same sentence.
- Vague 'this': Always clarify what "this" refers to
- Filler phrases: "In order to" -> "To"; "Due to the fact that" -> "Because"; "It is important to note that" / "It's worth noting" -> cut entirely; "Note that" / "Observe that" -> cut, just state the observation; "Try to X" -> "X"
- Excessive hedging: "It could potentially possibly be argued that X might have some effect" -> "X may affect Y". Limit "may" and "can" — hedge words should almost always be dropped. Either commit to the claim or cut it. Specific AI hedge phrases to strip: "arguably", "it could be said", "in many ways", "to some extent".
- Sycophantic tone: "Great question!", "You're absolutely right!", "I hope this helps!" -- chatbot residue; delete on sight
- Bare comparatives: "improves performance", "is more efficient" — compared to what? Always specify both sides of a comparison.
- Scare quotes: Do not use quotation marks to smuggle imprecise words in. If a term needs scare quotes, find a precise term instead.
- Unnecessary sentences: Ask of every sentence: "Is this necessary? Can I phrase this more simply?" Cut sentences that add no information.
- "Etc." and "and so on": Be exhaustive or pick representative examples. Trailing off signals lazy thinking.
- Number threshold: Write whole numbers below ten as words; use numerals for 10 and above. Be consistent throughout the document.
- Vague praise: "interesting", "fascinating", "groundbreaking" — without argument, these are empty. Explain why a result matters: first result, best performance, or new knowledge.
AI Writing Patterns to Avoid
These patterns signal machine-generated text. Never use them.
Banned Punctuation & Syntax
- Excessive em dashes — zero to two in the whole piece. A paired parenthetical (
-- like this --) counts as one use, and -- in markdown source counts the same as a typeset —. Replace with periods, commas, parentheses, or cut entirely. Not a rhythm tool. Prefer ;.
- "It's not X, it's Y" — formulaic contrast
- "You're not X, you're Y" — formulaic contrast
- "No X. No Y. Just Z." — rule-of-threes literary variant
- "An X with Y and Z" — dismissive constructions
- Mid-sentence self-questioning — "And honestly? That's amazing."
- doesn't just — "doesn't just" is a cliche
- Copula avoidance — "serves as", "stands as", "marks", "represents", "boasts", "features", "offers" instead of "is"/"are"/"has". Use the copula.
Banned Vocabulary
| Never use | Why |
|---|
delve | 2,700% spike post-ChatGPT |
tapestry, woven | False complexity signals |
intricate, interplay | Same |
underscore, highlight, showcase | AI emphasis verbs |
meticulous, adept, swift | Precision/speed cliches |
navigate (metaphorical) | "Navigate challenges" |
landscape (metaphorical) | "The AI landscape" |
robust | Meaningless intensifier |
leverage (verb) | Corporate AI-speak |
nuanced | Usually isn't |
consult | Corporate AI-speak |
shaped by, shaped this | Corporate AI-speak |
additionally, furthermore, moreover, ultimately, indeed, that said, in conclusion | AI connective tissue |
crucial, pivotal, vital, plays a crucial role | Inflated importance |
ecosystem, synergy, the space (vague nouns) | Replace with concrete things |
realm, dive into | AI spatial metaphors |
meaningfully, seamless, credible, beautifully unpredictable, quietly radical, disproportionate effect | AI modifier tells |
enduring, lasting | Legacy puffery |
foster, cultivate | AI growth verbs |
enhance | Vague improvement |
garner | AI synonym for "get" |
vibrant, rich (figurative) | Promotional tone |
nestled, breathtaking, stunning | Travel-brochure AI |
testament | "stands as a testament to" |
align with | Corporate AI-speak |
profound | Almost never earned |
interesting, fascinating | Empty praise — explain why it matters |
groundbreaking | Let the reader judge significance |
complex, rich (as praise) | Vacuous; describe what makes it so |
spine (metaphorical) | "the spine of the argument" — AI structural metaphor |
Banned Atmospheric Words
- ghosts, shadows, whispers, echoes (spectral obsession)
- "liminal" — overused atmosphere word
- forced quietness — "soft hum of distant conversation" at a loud party
- synesthesia abuse — "the texture of embarrassment," "Thursday tastes of almost-Friday"
Banned Rhetorical Moves
- Compulsive tricolons — a balanced three-item list ("X, Y, and Z") is a fingerprint. Break to two, split across sentences, or disrupt with an off fourth item. The ban targets balanced abstract triads; a list of concrete specifics ("blog posts, conference slides, and LinkedIn thought pieces") earns its place, and a four-item list is already the fix, not the disease.
- Formulaic openings — "In today's fast-paced world", "In the age of X", "There is something almost mythic about..."
- Empty profundity — "carve your code into my core, etched like prophecy"
- Mixed metaphors — piling concepts until collapse
- Sensory abstractions — attaching physical senses to abstract concepts
- Synonym cycling — calling the same thing "the protagonist", "the main character", "the central figure", "the hero" across consecutive sentences. Pick one name and stick with it.
- False ranges — "from X to Y, from A to B" where X/Y aren't on a meaningful scale. "From the Big Bang to dark matter" is not a range.
Content Inflation Patterns
- Significance puffery — "marking a pivotal moment in the evolution of...", "underscoring its vital role in..." Strip these. State what happened; let the reader judge importance.
- Vague attribution — "Experts believe", "Industry observers note", "Several sources suggest". Name the source or cut the claim.
- Formulaic challenges sections — "Despite challenges... continues to thrive." State the specific problem and what was done about it.
- Generic positive conclusions — "The future looks bright", "Exciting times lie ahead." End with a concrete fact, not optimism.
Human Signals — sprinkle these in
Small imperfections that make prose sound like a person. From the Pangram reflection convergence.
- Real asides — "give or take", "or whatever", "I forget the exact number", "this took me longer than I'd like to admit"
- Concrete specifics — brand names, dollar amounts, street names, weather, times ("8 a.m."), named objects. Replaces vague abstract nouns.
- Vary list/step openings — in numbered or stepped content, don't make every item start with an imperative verb. Alternate structures.
Paragraph Craft
- Lead and end with strong sentences — The first sentence of a paragraph states the point. The last sentence drives it home. Middle sentences elaborate.
- No orphan words — A single word alone on the last line of a paragraph wastes space and looks ugly. Shorten a sentence in the paragraph to fix the layout.
- Explain uncommon terminology on first use — Define jargon, acronyms, or domain-specific terms the first time they appear.
Academic & Conference Papers
These rules apply to formal academic writing. They override the conversational register elsewhere in this guide.
General
- Expand contractions — "it's" -> "it is", "don't" -> "do not"
- Unfold possessive apostrophes — "the model's accuracy" -> "the accuracy of the model". Formal register; aids non-native readers.
- Do not start every sentence with "We" — Vary sentence openings. "We train... We evaluate... We find..." is monotonous.
- Do not begin sentences with conjunctions — No "And", "But", or "Or" at the start of a sentence in formal writing.
- Minimum 3 sentences per paragraph — A 1-2 sentence paragraph signals an underdeveloped idea. Occasional exceptions for transitions.
- Describe what your method does, not what it doesn't — Positive framing. "Our method avoids X, Y, Z" tells the reader nothing about what the method is.
- No hostages to fortune — Avoid claims vulnerable to easy disagreement. Replace absolute claims with qualified ones ("many" instead of "most"). If you cannot defend a one-line boast, cut it.
- Run a spell checker before final submission — Overleaf misses errors that dedicated tools (e.g. Grammarly, LanguageTool) catch.
Abstract
Structure the abstract as a 2-minute spotlight talk:
- Contextualise the problem (1-2 sentences)
- Identify the gap in existing approaches (1 sentence)
- State the contribution (1 sentence)
- Key results with concrete numbers (2-3 sentences)
Include quantitative results directly — do not tease them. Avoid generic openings that could apply to any ML paper.
Introduction
- Arrive at the contribution quickly — Lengthy front-matter bores reviewers. State what you did within the first page.
- Lead with a compelling real-world example — Then formalise the abstract problem. Close the loop by addressing the motivating case in experiments.
- Anticipate critical questions — Answer them before the reader raises them.
Layout
- Minimise white space — Dense layouts let you fit more content in page-limited submissions. Applies to figures, captions, section headers, and paragraph spacing.
- Eye-catching first-page figure — Most readers decide whether to continue based on the first page. A clear, compelling figure earns their attention.
- Spend writing time in proportion to reading time — Title, abstract, and introduction receive the most reader attention. Spend equal effort on each. (Adapted from Jitendra Malik.)
- Balanced sections — Section titles should belong to the same scope. A section should contain more than one subsection.
Citations
- Citation grammar — A parenthetical citation must be removable without breaking the sentence. "Wilson et al. (2016) showed..." not "(Wilson et al. 2016) showed..."
- Cite generously — Especially work by likely reviewers. Fill the references section; blank bibliography pages signal carelessness.
- Cite throughout the paper — Not only in the Related Work section. Recent work (last 5-10 years) deserves inline citations.
Figure Captions
Every caption should answer three questions in order: what am I looking at, how should I read it, and what should I conclude.
- Self-contained — A reader should understand the figure without reading the surrounding text. You could rip any figure out and it would still make sense.
- Define terms inline — Don't assume the reader remembers what a metric means. Restate definitions in the caption (e.g. "0 = uniform prediction, 100 = perfect foresight"; "replaceability = reduction in score when removing a model").
- State the takeaway — Don't just describe axes and labels. Say what the reader should conclude. "gpt-oss-120b climbs 7 points and marginally exceeds Gemini 3 Pro" not "shows training curves". State it directly, not as an instruction: "The crosshatching makes dots hard to count", never "Note the crosshatching". Captions are where didactic "Note the X" openers sneak in.
- Include specific numbers — Ensemble weights (40%, 20%, 20%, 20%), score improvements (38.6 → 45.8), chosen baselines (18.8%). Ground the caption so the figure is verifiable.
- Explain design choices — Why the Y-axis is truncated at 22.3, what marker colour encodes, why a particular baseline is chosen. Prevent misreading.
- Short enough to scan — Most captions should be 2–3 sentences. Earn extra length only by packing in necessary context.
- 1-3 lines — Avoid paragraph-length captions. The main text carries the argument.
- State direction — Clarify whether higher or lower is better when the answer is ambiguous.
Essay Style
- Prose should be elegant and essayistic, with a tone that feels like a particularly thoughtful friend talking you through an idea.
- Use well-chosen concrete examples. Illustrate a point about the anxiety of status by describing the experience of walking through an airport or attending a dinner party. This grounds abstract ideas in recognisable emotional moments.
- Do not ask readers to accept an abstract principle on its own terms.
- Write the way a brilliant after-dinner speaker talks: digressive and funny, happy to make a bold claim and then back it up with a surprising example rather than a footnote.
- Reframe — take a familiar experience and make the reader see it differently. The reader should take pleasure in the shift in perspective.
- Write with genuine warmth towards human fallibility. Do not be cynical. Find contradictions and confusion endearing rather than contemptible. Make the reader feel understood rather than judged.
- Prefer unfolding cadence to drumbeat cadence. Let a paragraph develop an idea, turn it once, then land it. Avoid stacks of short verdict sentences unless you need a deliberate jolt.
- Do not make every insight sound like a slogan. A line can be neat; it should not feel pre-fabricated. If a sentence reads like a maxim dropped in from above, soften it with explanation, example, or a human observation.
- Let conclusions feel discovered, not announced. The final instruction or claim should sound like the natural end of the argument, not a poster slogan pasted on top.
- Mix sentence lengths. A short sentence works best after a longer one has set it up. Too many short declarative sentences in a row make the prose sound abrupt and prosecutorial.
- Preserve warmth in analytical writing. When describing confusion, ambiguity, or disagreement, sound interested rather than exasperated.
Technical Writing
- Concrete numbers beat vague claims. "2% error with 1.5KB" not "very efficient."
- One good analogy is worth five paragraphs of explanation.
- No throat-clearing ("In this post we will explore..."). Just start.
- No "In conclusion", "Let's dive in", or "Consider".
Exemplar Quotes
These show the target register — witty, concrete, self-aware:
- "The difference between hope and despair is a different way of telling stories from the same facts"
- "Anyone who isn't embarrassed of who they were last year probably isn't learning enough"
- "The human mind does not run on logic any more than a horse runs on petrol."
- "A flower is simply a weed with an advertising budget."
- "It is much easier to be fired for being illogical than it is for being unimaginative."
- "It is better to be vaguely right than precisely wrong."
- "Most of our childhood is stored not in photos, but in certain biscuits, lights of day, smells, textures of carpet."
A Note on Rules
This guide contains internal tensions. "Theme or rheme" is jargon in a guide that bans jargon.
Style guides are heuristics, not laws. Break any rule when following it would make the writing worse — Orwell's sixth rule applies to the whole document. The point is awareness: know when you're breaking a rule and why.
When editing AI-generated or AI-assisted text, do a final pass: ask "What still reads as obviously machine-generated?" Fix those tells, then check once more.
When editing, protect informality and personality — but nothing else. Do not:
- Standardise punctuation that carries tone — a casual hyphen-dash or exclamation mark often signals self-aware humour. "even if I know a few!" is wry; "even if I know a few." is flat.
- Remove personality for consistency — mechanical uniformity kills voice. A deliberate exclamation or informal dash is not an error.
The author has a penchant for semi-colons; leave them be.
This protection is never an exemption from the content rules. A slogan, an unglossed metaphor pair, or a having-it-both-ways verdict fails this guide whether it is new text or carried over verbatim from an earlier draft — "it was already there" and "that's the author's framing" are not defences. Pre-existing prose gets the same pass as new prose. (This clause exists because a carried-over maxim — "survived as structure, failed as arithmetic" — was once waved through as voice; it was spin.)
Ask: does this change make the writing better, or just more uniform?
Scripts
Two helpers ship in this skill's directory:
prose-lint.sh <file ...> — scans files for the patterns above (banned vocabulary, copula avoidance, filler, sycophancy, significance puffery, negative parallelisms, unnecessary words and hedges, superficial "-ing" openers, vague attribution, rule-of-three). Word lists are kept in sync with this guide. Skips markdown tables, import lines, and YAML frontmatter keys. Requires rg; exits non-zero when it finds hits.sentence-stats.sh <file> — per-sentence word counts plus a summary (average, range, standard deviation). Flags sentences over 40 words, low rhythm variety (stddev < 4), and a lack of short punchy sentences. Strips YAML frontmatter, tables, imports, and JSX before counting, so MDX files report prose only.
Run both before the manual editing pass; treat hits as candidates to review, not automatic rewrites.
References
