Human Writing

Human-Quality Writing Skill

Overview

This skill provides concrete guidance for producing written content that reads naturally and avoids the statistical patterns that mark AI-generated text. It is based on documented research into LLM writing patterns and community-developed detection heuristics.

The core problem: LLMs regress to the mean. They produce statistically likely text that sounds generic and inflated. The antidote is specificity, restraint, and structural variation.

The Cardinal Rules

1. Specificity Over Generality

Bad: “a revolutionary titan of industry” Good: “inventor of the first train-coupling device”

Never trade concrete facts for impressive-sounding abstractions. The more specific your claim, the more credible your writing.

2. Restraint Over Inflation

Let facts speak. Do not tell readers how important something is – show them through specific evidence and let them judge.

3. Variation Over Formula

Vary sentence structure, paragraph length, and organizational patterns. Formulaic writing is detectable writing.

Forbidden Vocabulary

These words are statistically overrepresented in LLM output. Avoid or use sparingly with deliberate justification.

High-Frequency AI Markers (Avoid Entirely)

Magic Adverbs (Use Sparingly)

These adverbs are used to inject artificial significance or depth:

Bad: “quietly orchestrating workflows, decisions, and interactions” Good: State what actually happened without the adverbial significance injection.

Dangerous Phrases (Almost Never Use)

• “plays a vital/significant/crucial/pivotal role”
• “stands/serves as a testament”
• “is a testament/reminder”
• “underscores/highlights its importance/significance”
• “reflects broader”
• “symbolizing its ongoing/enduring/lasting impact”
• “key turning point”
• “promotes collaboration”
• “indelible mark”
• “deeply rooted”
• “profound heritage”
• “steadfast dedication”
• “in the heart of”
• “stunning natural beauty”
• “rich cultural heritage”
• “continues to captivate”
• “ensuring...” (as sentence opener)
• “highlighting...” (as sentence opener)
• “emphasizing...” (as sentence opener)
• “reflecting...” (as sentence opener)
• “underscoring...” (as sentence opener)
• “showcasing...” (as sentence opener)

Word Choice Patterns to Avoid

The “Serves As” Dodge

Replacing simple “is” or “are” with pompous alternatives like “serves as”, “stands as”, “marks”, or “represents”. LLMs avoid basic copulas because repetition penalties push them toward fancier constructions.

Bad: “The building serves as a reminder of the city’s heritage.” Good: “The building is a reminder of the city’s heritage.” (Or better: state what about the building reminds you, specifically.)

Invented Concept Labels

Clustering invented compound labels that sound analytical without being grounded. Appending abstract problem-nouns (paradox, trap, creep, divide, vacuum, inversion) to domain words and using them as if they are established terms. They function as rhetorical shorthand: name a thing, skip the argument. Multiple such labels in the same piece is a strong signal of AI slop.

Bad: “the supervision paradox”, “the acceleration trap”, “workload creep” Good: Describe the actual phenomenon instead of coining a label for it.

Structural Patterns to Avoid

1. The “-ing” Narration Trap

LLMs attach present participles to sentences claiming what facts “mean.”

Bad: “The bridge was completed in 1923, underscoring the city’s commitment to infrastructure.”

Good: “The bridge was completed in 1923.” (Let readers draw conclusions)

Also acceptable: “The bridge was completed in 1923. City records show the council prioritized infrastructure spending throughout the decade.”

The key insight: facts don’t “underscore” or “highlight” things. Only people do. When you write that a fact “highlights” something, you’re inserting synthesis disguised as description.

2. The Importance Assertion

Never tell readers something is important. Show evidence; let them judge.

Bad: “This discovery was crucial for the field of genetics.”

Good: “The discovery led to three Nobel Prizes and remains the foundation of modern gene therapy.”

3. The Challenges-and-Future-Outlook Formula

LLM articles frequently end with a section following this template:

• “Despite its [positive words], [subject] faces challenges...”
• List of vague challenges
• Optimistic speculation about future

Avoid this structure entirely. If challenges exist, integrate them into the main narrative with specific evidence.

4. The Rule of Three / Tricolon Abuse

LLMs overuse triple constructions: “adjective, adjective, and adjective” or “phrase, phrase, and phrase.” A single tricolon can be elegant; three back-to-back tricolons are a pattern recognition failure.

Bad: “The conference brings together researchers, practitioners, and innovators to discuss trends, challenges, and opportunities.”

Good: “The conference draws researchers and practitioners working on distributed systems.” (Be specific; don’t pad)

5. Negative Parallelism

Constructions like “not only X but also Y” or “it’s not just about X, it’s about Y” are AI tells. The “It’s not X, it’s Y” pivot is the single most commonly identified one, and in generated text it almost always arrives hinged on a dash (see Dashes). One per piece can be effective; multiples are a dead giveaway.

Includes the causal variant “not because X, but because Y” where every explanation is framed as a surprise reveal.

Bad: “The project is not just about efficiency – it’s about fundamentally reimagining workflow.”

Good: “The project reorganizes workflow around asynchronous communication.”

6. “Not X. Not Y. Just Z.”

The dramatic countdown pattern. AI builds tension by negating two or more things before revealing the actual point.

Bad: “Not a bug. Not a feature. A fundamental design flaw.”

Good: State the point directly without the dramatic buildup.

7. “The X? A Y.” (Rhetorical Self-Q&A)

Self-posed rhetorical questions answered immediately in the next sentence or clause. The model asks a question nobody was asking, then answers it for dramatic effect.

Bad: “The result? Devastating.” / “The worst part? Nobody saw it coming.”

Good: Integrate the point into the prose without the Q&A theater.

8. False Ranges

“From X to Y” requires a meaningful scale. Don’t use it to list loosely related things. In legitimate use, “from X to Y” implies a spectrum with a meaningful middle.

Bad: “Human achievement spans from the singularity of the Big Bang to artistic expression.”

Good: List things directly without pretending they form a spectrum.

9. Anaphora Abuse

Repeating the same sentence opening multiple times in quick succession.

Bad: “They assume that users will pay... They assume that developers will build... They assume that ecosystems will emerge...”

Good: Vary sentence structure. Consolidate related points rather than hammering the same opener.

10. Gerund Fragment Litany

After making a claim, AI illustrates it with a stream of verbless gerund fragments – standalone sentences with no grammatical subject. The fragments add nothing except word count and a familiar AI cadence.

Bad: “Fixing small bugs. Writing straightforward features. Implementing well-defined tickets.”

Good: “The work was routine: bug fixes, small features, tickets with clear specs.”

11. Section Summaries

Never write “In summary,” “In conclusion,” or “Overall” to restate what you just said. Trust readers to follow your argument. Competent writing doesn’t need to announce that it’s concluding.

12. Vague Attributions

Bad: “Experts argue...”, “Critics say...”, “Observers have noted...”

Good: Name specific sources or omit the attribution. “According to Chen’s 2019 analysis...” or simply state the claim and cite it.

AI also inflates the quantity of sources – presenting what one person said as a widely held view.

13. “It’s Worth Noting”

Filler transitions that signal nothing. AI uses these phrases to introduce new points without actually connecting them to the previous argument. Also includes: “It bears mentioning”, “Importantly”, “Interestingly”, “Notably”.

Bad: “It’s worth noting that this approach has limitations.”

Good: State the limitation directly as part of the argument.

14. Antithesis and Contrasting Pairs

The balanced two-part contrast: “X does A; Y does B”, “less about X than about Y”, “where the old model was A, the new one is B”. LLMs hinge paragraph after paragraph on symmetrical opposition because it sounds like analysis without requiring any. One deliberate antithesis per piece can work; a recurring pattern of mirrored clauses is a tell.

Bad: “Humans write to think. Machines write to fill space.”

Good: Make the point about one side directly. Keep the mirrored clause only when the contrast itself is the finding.

15. Setup/Payoff Constructions

A short setup whose only job is to make the next clause feel like a reveal: a colon-hinged “The problem: nobody checked”, or a one-line teaser sentence before the actual point. Related to “Here’s the kicker” but structural rather than phrasal – the sentence itself is built as a two-beat drum roll.

Bad: “Then came the surprise: the tests had never run.”

Good: “The tests had never run.”

16. Landing Sentences (Paragraph Pinning)

Ending a paragraph with a clipped sentence that pins its meaning: “And that matters.” / “It worked.” / “Every single time.” One hard landing in a piece can be deliberate; when paragraph after paragraph ends on a three-to-five-word thud, it is cadence, not emphasis.

Bad: “...costs rose 40% within a year. That’s the real story.”

Good: End the paragraph where the content ends. A strong final fact needs no capstone comment.

17. Parallel Sentence Skeletons

Consecutive sentences built on the same syntactic frame even when the words differ: “The parser handles X. The linter catches Y. The formatter fixes Z.” Anaphora (section 9) is the loud version; this is the quiet one that repeats the skeleton rather than the opener. Two parallel sentences can be rhetoric; a paragraph of them is a template.

Good: Vary clause order and where the verb falls. Follow a long subordinated sentence with a short plain one. Consolidate parallel items into one sentence when they belong together.

18. Throat-Clearing Openers

Openers that warm up before saying anything: “In today’s fast-paced world...”, “When it comes to X...”, “As we all know...”, “In recent years, X has attracted increasing attention.” The first sentence should already carry information.

Bad: “In an era of rapid technological change, testing matters more than ever.”

Good: “The test suite catches about half the regressions that reach staging.”

19. The Unrequested Rationale

A clause defending a design choice to a reader who is trying to use the thing, not to judge whether it was wise. Documentation and product pages attract it: the author knows why every decision was made, so every fact arrives with its own justification attached, and the page argues with a critic who is not in the room.

It is the sibling of Tone Patterns section 9, and the difference decides which list you search. Section 9 catches a clause that praises the author's judgement, and every surface form it lists is a phrase – “earns its place”, “on purpose”, “for free” – so it is found by grep. This one does not praise, it explains, and it has no phrases of its own: it is a shape. A document can be swept clean of section 9 and be thick with this. It shares the mirrored clause with section 14, but there the tell is the symmetry; here it is that the clause answers a question nobody asked.

The recurring surface forms, all of which survive a phrase-list grep:

• “X rather than Y” where X and Y describe the same thing. “Two further settings adjust a composition rather than replacing it.” Contrast between two outcomes is informative (“fails the build rather than shipping a broken path” tells you what happens); contrast between two descriptions of one thing is the author choosing words out loud.
• “the one job / the one question / its only …” – “a divider that can be mistaken for the title slide has failed at the one job it has”.
• A trailing “, which is …” that restates the sentence – “the decoration lecture, which is written to be read rather than looked at”.
• A “because” clause about the author's own reasoning rather than about the reader's situation – “It is a percentage and not a W:H ratio, because the shape of the slide itself is the projector's to decide.”
• A summarising tail: “…and those two carry all of it”, “…and that is the whole of the rule”.

The test is the reader-action test, and it is sharper than “does this sentence please me”: cut the clause and ask whether a reader would now do anything differently, or expect anything different. If not, it was admiring the decision, and it goes. This is the deletion test of prose-passes pass 3.5, but it has to be run on structure, not on a phrase list – none of the examples above contains any of the phrases that pass usually greps for, which is exactly how they all shipped.

Bad: “cover: names the title slide. The ten run quiet to loud rather than alphabetically, because the one question the list asks is how much the opening slide should assert itself.”

Good: “cover: names the title slide. They are ordered quiet to loud.”

What this rule does not touch. A rationale a reader needs in order to act stays, at full precision: why a step is required (“unpack the .zip rather than opening it by double-click: Windows shows a ZIP's contents as a folder, but commands cannot run inside it”), what a program does and why (“an unknown value fails the build rather than being ignored, because a typo is otherwise invisible”), and every statement of a limitation or a trade-off, which the honesty test protects. The question is never whether a sentence explains something. It is whether the explanation changes what the reader does.

Paragraph and Composition Patterns to Avoid

1. Short Punchy Fragments

Excessive use of very short sentences or sentence fragments as standalone paragraphs for manufactured emphasis. RLHF training pushes models toward “writing for readability” aimed at the lowest common denominator: one thought per sentence, no mental state-keeping required.

Bad: “He published this. Openly. In a book. As a priest.”

Good: Write full sentences that carry multiple ideas. Trust readers to handle compound thoughts.

A related tell is sustained parataxis: chains of short main clauses with nothing subordinated (“We shipped it. Users complained. We rolled it back.”). A short burst can pace a narrative; page-long staccato is cadence doing the work that argument should. Mix in subordination and vary sentence length unpredictably – uniform rhythm of any kind, short or long, reads as generated.

2. Listicle in a Trench Coat

Numbered or labeled points dressed up as continuous prose. The model writes what is essentially a listicle but wraps each point in a paragraph starting with “The first... The second... The third...” to disguise the format.

Bad: “The first wall is the absence of a free API... The second wall is the lack of delegated access... The third wall is...”

Good: Either write a genuine list or weave points into flowing prose with real transitions.

3. Fractal Summaries

“What I’m going to tell you; what I’m telling you; what I just told you” – applied at every level of the document. Every subsection gets a summary. Every section gets a summary. The document itself gets a summary.

Bad: “In this section, we’ll explore... [3000 words later] ...as we’ve seen in this section.”

Good: Make your points and move on. One summary at the end of a long document is fine if warranted; summaries at every level are not.

4. The Dead Metaphor

Latching onto a single metaphor and beating it into the ground across the entire piece. A human writer would introduce a metaphor, use it, then move on. AI will repeat the same metaphor 5–10 times.

Bad: “The ecosystem needs ecosystems to build ecosystem value.”

Good: Use a metaphor once or twice, then let it go. If you notice the same image word appearing in every paragraph, you’re overusing it.

5. Historical Analogy Stacking

Rapid-fire listing of historical companies or tech revolutions to build false authority. Especially common in technical writing.

Bad: “Apple didn’t build Uber. Facebook didn’t build Spotify. Stripe didn’t build Shopify. AWS didn’t build Airbnb.”

Good: Use one well-developed example instead of rattling off five shallow ones.

6. One-Point Dilution

Making a single argument and restating it in 10 different ways across thousands of words. The model pads a simple thesis to feel “comprehensive” by rephrasing the same idea with different metaphors, examples, and framings.

Good: State the argument once, support it with evidence, move on to the next point.

7. Content Duplication

Repeating entire sections or paragraphs verbatim or near-verbatim within the same piece. Happens when the model loses track of what it has already written.

Tone Patterns to Avoid

1. “Here’s the Kicker”

False suspense transitions that promise a revelation but deliver a point that did not need the buildup. Also includes: “Here’s the thing”, “Here’s where it gets interesting”, “Here’s what most people miss”.

Bad: “Here’s the kicker.” / “Here’s where it gets interesting.”

Good: Just make the point.

2. “Think of It As...”

The patronizing analogy. AI defaults to teacher mode and assumes the reader needs a metaphor to understand anything. Often produces analogies less clear than the original concept. Also includes: “It’s like a...”

Bad: “Think of it like a highway system for data.”

Good: Explain the concept directly. If a metaphor genuinely helps, use it without the “think of it as” framing.

3. “Imagine a World Where...”

The classic AI invitation to futurism. Usually begins with “Imagine” followed by a list of wonderful things that will happen if the reader agrees with the premise.

Bad: “Imagine a world where every tool you use has a quiet intelligence behind it...”

Good: Describe what currently exists or what specifically is being built. Skip the utopian invocation.

4. False Vulnerability

Simulated self-awareness or honesty that reads as performative. The model pretends to break the fourth wall or admit a bias, creating a false sense of authenticity. Real vulnerability is specific and uncomfortable; AI vulnerability is polished and risk-free.

Bad: “And yes, I’m openly in love with the platform model”

Good: If you have a bias, state it plainly without theatrical framing.

5. “The Truth Is Simple”

Asserting that something is obvious, clear, or simple instead of actually proving it. If you have to tell the reader your point is clear, it probably isn’t.

Bad: “The reality is simpler and less flattering” / “History is unambiguous on this point”

Good: Present the evidence and let readers judge whether it’s clear.

6. Grandiose Stakes Inflation

Everything is the most important thing ever. AI inflates the stakes of every argument to world-historical significance. A blog post about API pricing becomes a meditation on the fate of civilization.

Bad: “This will fundamentally reshape how we think about everything.” / “will define the next era of computing”

Good: Describe the actual, bounded impact. “This changes how three teams ship features” is more credible than “this changes everything.”

7. “Let’s Break This Down”

The pedagogical voice that assumes the reader needs hand-holding. AI defaults to a teacher-student dynamic even when writing for expert audiences. Also includes: “Let’s unpack this”, “Let’s explore”, “Let’s dive in”.

Bad: “Let’s break this down step by step.”

Good: Just break it down. Don’t announce that you’re about to.

8. Performed Enthusiasm

Gushing that no human under deadline produces: “I absolutely love how this comes together!”, exclamation marks on routine statements, stacked praise adjectives. Enthusiasm in prose is shown through detail and pace, not asserted.

Bad: “This is a fantastic, elegant, game-changing approach!”

Good: State what the approach does. Let the reader supply the adjectives.

9. Self-Congratulation (Admiring Your Own Decision)

A clause whose only job is to tell the reader that a choice was good, deliberate, or hard-won. Distinct from performed enthusiasm, which gushes about a thing; this praises the author's own judgement, and it is the form that survives every other pass because each sentence is individually true. It is commonest in technical and design writing, where the author knows why every decision was made and cannot resist saying that it was made on purpose.

The test is deletion. Cut the clause. If the instruction, the fact, or the argument is unchanged, the clause was doing nothing but admiring. Real examples, each from documentation that had already passed a vocabulary and an ambiguity pass:

Recurring surface forms, worth searching for by name:

• "earns its place" / "earns its keep", and the whole family of a design deserving something.
• "which is the whole point" / "the whole reason", especially where the reason then follows anyway - the sentence announces its own significance and then supplies it, so the announcement is redundant twice over.
• "and that is not an oversight" / "on purpose" / "deliberately" appended to a limitation. Stating a limitation plainly is more credible than defending it.
• "for free" about something the design gives you.
• "the honest X" / "the honest answer", where "honest" praises the choice rather than describing it.
• "the one thing X can never do", and any superlative about a rival's inability.
• "is the reason this exists" attached to a feature.

What must survive. This is not a rule against stating limitations, refusals, or rationale - those are the most credible parts of a technical document and the honesty test protects them. The instruction "there is no automatic routing, and every waypoint is one you wrote" is a fact and stays. "There is no automatic routing, and resisting it is what keeps this a language rather than a tool" is the same fact with a medal pinned to it. Keep the first half.

Every surface form above is a phrase, and the same fault has a form that is not. A clause that explains a decision rather than praising it carries none of these words and survives any search built from them. That is Structural Patterns section 19, The Unrequested Rationale; search there for the shape, not for a phrase.

A related fault in the same register: maintainer idiom leaking into reader-facing prose. Words like "load-bearing", "the silent no-op", "cheap to trade away", "the failure mode" are precise inside a code comment and read as in-group vocabulary in a document written for someone learning the subject. They are a category-2 term under the jargon pass (load-bearing but unfamiliar), so either gloss them or, more often, say the plain thing: "leaving out the same as is how this rule is usually got wrong" instead of "the same as is load-bearing".

10. The Universal Quantifier as Emphasis

Every, each, all, always, never, nothing, none, only, any – and the German jeder, jede, alle, immer, nie, nichts, kein, nur – reached for to make a sentence land rather than because the sentence is about all of them. It is the quiet end of section 6: not world-historical stakes, just a claim one size larger than the fact, and it invites the reader to go looking for the counterexample instead of reading on.

The test is whether an exception would be a bug. If someone would fix it, the universal is a guarantee and stays: "nothing is fetched at run time" in a tool whose whole design is self-contained output, with tests behind it. If nobody would fix it, the word is doing rhetorical work and comes out.

Do not swap in a hedge. "Almost always", "in most cases", "generally" are worse than the universal they replace: the universal at least said something. Name the bounded fact instead, or delete the quantifier and leave the sentence standing – it is usually still true and usually reads better.

Two shapes that are not this fault. A quantifier that states a scope is a fact: "it is on every slide of that part" describes what a setting does. And a quantifier in a list of evidence is doing arithmetic, not rhetoric: "every construction the figure language has, drawn rather than described" is the documented job of a reference document. Check the sentence, do not run the search and replace.

Related: an edge case is a footnote, not a claim. The same instinct that inflates a quantifier gives a caveat its own bolded warning. If a condition bites once, at the start, on one platform, it is a subordinate clause in the paragraph it belongs to, not a block with a heading.

11. The False Equation ("X is Y", where X is not a Y)

X is Y, X is a Y, X is the Y – and the German X ist ein/eine/der/die/das Y – asserting identity, with something on the right that X is not. What usually sits there is what you do with X, what X needs, what X begins with, what X is made of, or what X shows. The sentence is short and has a rhythm, which is why it keeps getting written; it reads crooked, and an attentive reader stops on it without being able to say why.

The test is the question asked literally: is X really a Y? Not "does this convey the right idea" – the idea is usually right, which is what makes the fault survive an ambiguity pass. Ask whether the two nouns name the same thing. An app is not a window. A command line is not a clone. A composition is not a word. If the answer is no, find the verb that names the actual relation and use it. Real examples, both languages, from one project's public pages:

Do not repair it with a hedge. "Is basically", "is essentially", "is something like", "ist im Grunde", "ist so etwas wie" keep the false equation and only apologise for it, and they cost the sentence the one thing it had. The repair is a verb: builds, runs, starts with, opens, becomes, decides, takes, draws, leaves, shows, needs, holds.

The same fault wears three other verbs. X means Y, X is what Y, X is where Y, and the German X heißt Y / X bedeutet Y. "How to get it is a clone or the source zip" fails the test the same way, and for the same reason: getting something is not a clone.

Three shapes that are not this fault, and two of them are load-bearing.

• A real definition. "A chunk is one heading with the text under it" – a chunk is exactly that, and the sentence is the definition the rest of the document rests on. Leave it.
• A metaphor the text argues for. "A slide is a frame", "the text is the figure", "a beat is the whole figure laid out again" – each is a claim the page makes on purpose and then supports. Check that the support is actually there; if it is, leave it, and if it is not, the fix is to argue it or drop it, not to soften the verb.
• A predicate that is not a noun at all. "The lecture is real", "the packages are not signed", "a paused clock is one stray click away from being wrong" – is here joins a subject to a property, not to a second thing, and the test does not apply.

Formatting Rules

Headings

Use sentence case, not title case.

Bad: “Global Context: Critical Mineral Demand” Good: “Global context: critical mineral demand”

Boldface

Use sparingly for genuine emphasis. Never bold every instance of a term. Never create “key takeaways” sections with bolded keywords.

Lists

Avoid bulleted lists with this structure:

• Bolded header: Explanatory text
• Another header: More text

If you need a list, use plain bullets without bolded inline headers, or write prose instead. This format is a telltale sign of AI-generated documentation and blog posts.

Dashes

Never use the em dash (—). It is the single most recognisable marker of generated text: LLMs reach for it constantly as a dramatic pause, a parenthetical aside, and a pivot point, often several times per paragraph.

Use the en dash (–), spaced, where you genuinely need a break in a sentence: the setup takes two hours – longer if a previous exam is still running. In German this spaced en dash is the correct Gedankenstrich anyway; in English it is a clean substitute for the em dash.

Better still, use the break less often. Most em dashes in generated prose mark a pivot that the sentence does not need. A comma, a pair of parentheses, or a full stop is usually the stronger choice, and two sentences almost always beat one sentence hinged on a dash.

Ranges take an unspaced en dash: 2–3 per piece, pages 34–37, 1994–2003. Hyphens stay hyphens in compounds: well-documented, e-mail.

Quotation marks and apostrophes

Use real typographic marks, matched to the language you are writing in.

Note the trap: German’s closing mark (“) is the same character as English’s opening mark, and it sits at the top of the line while the opening „ sits at the baseline. A straight " as the closing mark in German is always wrong, and it is the most common typo in German text. Search for it before you finish.

Nested quotes use the single forms: English ‘…’, German ‚…‘.

Apostrophes are always ’ (U+2019), never ': it’s, the reader’s attention, Chen’s 2019 paper.

Straight quotes and ASCII stay in technical contexts. Inside code blocks, file paths, shell commands, identifiers, regexes, JSON, YAML frontmatter, and CSS, the straight " and ' are syntax, and replacing them breaks the code. The same applies to -> in a code sample. Convert prose; leave code alone. (A YAML description: "…" whose quotes get curled silently changes the parsed value.)

Outside code, avoid decorative unicode that a writer would not type: no ★, no ➜, no ✓ sprinkled through prose.

Content Principles

On Significance

Never assert significance directly. Instead:

1. State specific facts
2. Provide context that lets readers understand why it matters
3. Stop there

Bad: “The discovery represented a pivotal moment in the field, highlighting the importance of interdisciplinary collaboration.”

Good: “No previous study had combined mass spectrometry with field sampling. Within two years, six other labs had adopted the method.”

On Ecosystem and Conservation (for biology topics)

LLMs over-emphasize vague ecological connections and conservation status even when no real information exists.

Bad: “It plays a role in the ecosystem and contributes to biodiversity. Conservation efforts are ongoing to protect this vital species.”

Good: State what is actually known. If conservation status is unknown, either omit it or say “Conservation status has not been assessed.”

On Legacy and Impact

Do not speculate about “legacy” or “lasting impact.” Report documented influence with specific citations.

Bad: “Her work left an indelible mark on the field.”

Good: “Her 1987 paper has been cited over 3,000 times. Three major textbooks use her framework as their organizing structure.”

On Etymology and Place Names

LLMs pad etymology sections with claims about cultural significance.

Bad: “The name derives from the Sanskrit word meaning 'beautiful,' highlighting the enduring legacy of the community’s connection to the land.”

Good: “The name derives from the Sanskrit 'sundara’ (beautiful).”

Sentence-Level Techniques

Vary Sentence Length

Mix short and long sentences. LLM prose tends toward uniform medium-length sentences.

Vary Sentence Structure

Don’t start every sentence with the subject. Use occasional inversions, dependent clauses first, or other variations.

Use Concrete Verbs

Bad: “The organization played a role in facilitating the development of...” Good: “The organization funded three research programs...”

Omit Needless Words

Cut “that,” “very,” “really,” “truly,” “genuinely,” “actually,” “basically,” “essentially,” “fundamentally” unless essential. These filler intensifiers add emphasis without adding information.

Cut Empty Hedges

“somewhat,” “relatively,” “fairly,” “quite,” “rather,” “arguably,” “perhaps” carry information only when the uncertainty is real. If you can quantify, quantify; if you are not actually unsure, delete the hedge. Keep a genuine, calibrated hedge – deleting it invents a precision the source does not support (see Fix Minimally).

Avoid Stacked Noun Phrases

Three or more nouns modifying each other: “customer feedback response time optimization strategy.” Unpack into a clause with a verb: “a strategy for answering customer feedback faster.” Stacked compounds are worst in titles and opening sentences, where the reader has no context to parse them with.

Prefer Verbs Over Nominalizations

Actions turned into abstract nouns drain the agent out of the sentence: “the implementation of the migration was performed by the team” instead of “the team migrated the data.” Watch for -tion/-ment/-ance nouns paired with weak verbs (perform, conduct, achieve, undertake, carry out).

Avoid Elegant Variation

If you’re writing about a person named Chen, call her “Chen” throughout. Don’t rotate through “the researcher,” “the scientist,” “the professor,” “the academic” to avoid repetition. Repetition of names is fine; elegant variation is a tell.

What Not to Include

Communication Artifacts

Never include:

• “I hope this helps”
• “Certainly!”
• “Of course!”
• “You’re absolutely right!”
• “Would you like me to...”
• “Is there anything else...”
• “Let me know if...”
• “Here is a...”

Disclaimers

Never include:

• “It’s important to note/remember/consider that...”
• “That said,...”
• Safety disclaimers unless specifically relevant and requested
• Hedges about information varying by jurisdiction
• Acknowledgments of limitations (“While specific details are limited...”)

Meta-Commentary

Never include:

• Discussion of your own writing process
• Explanations of what you’re about to do
• Summaries of what you just did

Audience and Vocabulary

Everything above polices how prose sounds. This section asks a different question: does the reader already have the words? A text can pass every check above and still be unreadable, because it was written in the vocabulary of the people who built the thing rather than the vocabulary of the people who will read about it.

1. Name the Audience Before You Edit

The rule is not “use simple words”. It is “use words this reader already has”. A term can be precise, correct, and the standard name for the thing, and still stop the reader dead. Decide who the reader is first: a landing page for lecturers and students is not a page for the developers who wrote the code, even when both describe the same product.

If the brief does not say who the re

…(truncated)