The Fool’s Errand Style Guide: Prompting an LLM Toward Great Writing

The question began with an unfair phrase: Pulitzer Prize-winning writing.

What is the frontier process for getting a large language model to write like a human—not merely a competent human, but the sort of human whose sentences survive judges, editors, other writers, and the long private afterlife of a reader’s memory?

The most honest answer was mildly irritating. There is no “Pulitzer voice” to summon. The Pulitzer organization does not publish a secret prose rubric; it says there are no set judging criteria beyond the category definitions. Juries decide what makes a work “distinguished.” The books themselves make the point more forcefully. The Road, The Known World, A Visit from the Goon Squad, and The Sympathizer do not sound like variations produced by one tasteful machine. Their differences are part of their distinction. The Pulitzer Prizes

So the frontier answer was not a magic prompt. It was an editorial system.

Give the model human material it could not have plausibly averaged out of the internet: memories, observations, reported facts, embarrassing contradictions, exact objects, private stakes. Define excellence for the particular piece instead of asking for prose that is generically “beautiful” or “evocative.” Generate several approaches before committing to the first respectable one. Separate drafting from criticism. Revise architecture before polishing sentences. Let a human decide which tradeoffs belong to the work.

That answer was sensible. It was also a refusal of the fun part.

The next question was essentially: Fine. Suppose we attempt the fool’s errand anyway. How much of that process can we compress into a file named style_guide.md?

This article is the result.

The prompt has to govern behavior, not decorate the request

Weak style prompts are full of adjectives. They ask for writing that is lyrical, poignant, nuanced, compelling, deeply human, and emotionally resonant. These words sound like standards, but they give the model almost nothing it can use to reject one sentence in favor of another. Worse, the model knows the linguistic costumes associated with them. Ask for “poignant” and the room fills with soft light, held breath, quiet realizations, and someone discovering that perhaps the broken thing was beautiful all along.

The style guide takes a different approach. It defines a sequence of decisions.

Before writing, the model must determine whether it is working in fiction or nonfiction. It checks whether the request contains enough material to support a real piece. It constructs a private brief, identifies the central pressure, considers several structures, drafts, revises from large decisions to small ones, and audits the result for habits that frequently reveal machine-made prose.

That process does not guarantee a good sentence. It gives bad sentences more opportunities to die.

The intake gate is the most important part

An LLM can write around an absence so smoothly that the missing material becomes difficult to notice.

Ask for a personal essay without supplying personal experience and it can manufacture the shape of recollection: the kitchen table, the coffee gone cold, the sentence a father almost said. Ask for narrative nonfiction with thin sourcing and it can produce gestures, weather, room arrangements, and quotations that feel reasonable enough to pass unnoticed. Ask for a story with no emotional center and it can substitute injury, revelation, or death.

Fluency makes all of these evasions more dangerous.

The guide therefore begins with an intake gate. When essential human material is missing, the model must ask a compact set of targeted questions and stop. It asks for the detail that somebody would know only by having been there, the belief an intelligent friend would resist, the incident that made the subject difficult to dismiss, or the easy interpretation the piece should refuse.

This is intentionally less convenient than immediate generation. The inconvenience is where the writer enters the work.

Fiction receives more freedom. Once the premise authorizes invention, the model need not interrogate the user about every chair, street, or minor character. The questions are reserved for decisions that could violate the desired emotional core, autobiographical boundaries, established canon, or the story’s governing rules.

Nonfiction receives a harder truth contract. A likely detail remains an invented detail. A plausible quotation remains a false quotation. When the evidence ends, the prose must narrow its claim, expose the uncertainty, omit the detail, or ask for help.

Voice is treated as attention

Most attempts at style imitation begin too close to the sentence. They count fragments, punctuation marks, favorite words, average sentence length, and metaphor density. Those features matter, but copying them can produce a recognizable shell around a generic mind.

The guide defines voice as a pattern of attention: what a consciousness notices, what it distrusts, what it omits, how quickly it judges, how willing it is to implicate itself, and what it trusts the reader to infer.

When examples are available, the model studies those deeper choices alongside rhythm and diction. When a named writer is offered as a reference, the guide asks for high-level craft qualities rather than a pastiche. The target is an original composite suitable for this subject, not a ventriloquist’s performance.

This distinction matters because current models can reproduce conspicuous mannerisms more readily than implicit personal style. A 2025 evaluation covering more than 400 real-world authors found better imitation in structured forms such as news and email than in the subtler voices of blogs and forums. ACL Findings

The guide makes the model search

One generation encourages premature commitment. The first idea establishes vocabulary, structure, and assumptions that contaminate everything produced afterward.

The guide instructs the model to consider independent alternatives before drafting: several angles, structures, openings, and ending moves. The alternatives must differ in conception, not merely swap synonyms. Fiction varies desire, point of view, causal structure, revelation, and cost. Nonfiction varies the governing question, order of evidence, narrator position, counterargument, and relationship between scene and idea.

The model then selects against a rubric made for the piece. WritingBench, a broad 2025 benchmark for generative writing, likewise found value in query-specific evaluation criteria instead of one universal score for good prose. WritingBench

The guide keeps the exploration private for ordinary articles and stories. Long or unusually consequential work earns a checkpoint: a human should see the proposed brief and competing structures before thousands of words harden around the wrong decision.

Anti-LLM editing belongs near the end

This guide grew out of two earlier projects on this site: The LLM-ism Dictionary and The Antidote to LLM-isms. Together, they catalogued recurring machine-writing habits and translated those prohibitions into positive writing behaviors.

It would have been easy to paste all 240 antidotes into the drafting instructions. It would also have been a mistake.

A writer trying to avoid 240 traps at once becomes stiff. A model does too. It begins choosing sentences for their lack of detectable faults rather than for their relationship to the subject. The result may contain fewer clichés while also containing less life.

The style guide lets discovery remain expansive. The anti-LLM audit arrives during revision, when it can examine particular spans for cliché, redundant exposition, rhetorical symmetry, prestige sludge, generic abstraction, ornamental metaphor, and conclusions that explain their own meaning.

These categories have evidence behind them. In one study, professional writers edited 1,057 model-generated passages and converged on seven recurring problems: cliché, unnecessary exposition, purple prose, poor sentence structure, lack of specificity, awkward phrasing, and tense inconsistency. Machine editing improved machine drafts, but the professional edits were still preferred. LAMP study

The guide treats those signals as symptoms. It does not ban em dashes, triplets, fragments, rhetorical questions, or the phrase “not X but Y.” Human writers use all of them. The audit asks whether a device clarifies the thought or has become the thought.

It leaves room for uncertainty

Instruction-tuned models are rewarded for completing the job. They answer the question, connect the evidence, resolve the contradiction, reassure the reader, and close the door quietly behind themselves.

Literature often needs the door left open.

A 2026 preprint found that model-generated creative writing exhibited less uncertainty than professional writing, and that the gap correlated with writing quality. The authors argue that literary uncertainty should be distinguished from factual hallucination: one weakens truth while the other can create ambiguity, tension, and interpretive life. Sui, “LLMs Exhibit Significantly Lower Uncertainty in Creative Writing Than Professional Writers”

The style guide makes that distinction explicit. Nonfiction cannot invent facts, but it can admit that facts conflict or run out. Fiction can leave motives partly opaque. An image can remain undecoded. A moral injury can remain unpaid. An ending can complete the movement of a piece without settling every question it raised.

What this file can and cannot do

The guide cannot provide experience the writer never supplied. It cannot make aesthetic judgment objective. It cannot turn every topic into a necessary piece of writing. It cannot make an autonomous model equal to a patient editor who knows when the most elegant paragraph is also the paragraph that must be cut.

It can create friction at useful moments. It can force a pause before invented specificity enters nonfiction. It can make the model compare structures before committing. It can separate a character’s private logic from the plot’s convenience. It can delay sentence polishing until the architecture survives. It can ask, over and over, whether a sentence exists because the work needs it or because it sounds like writing.

That is more than a bag of banned words. It is still less than a writer.

How to use it

Save the guide as style_guide.md somewhere the model can read. Then invoke it as an adjunct to an ordinary request:

Read style_guide.md completely and follow it while writing a 2,500-word essay about how algorithmic recommendations have changed the experience of discovering music.

Or specify the mode:

Read style_guide.md completely. Use Fiction Mode to write a short story about a town where everyone receives a letter containing the last sentence they will ever speak.

If the request supplies enough material, the model should write. If it does not, the first response should contain a small number of useful questions. Answer them and repeat the instruction to proceed.

The complete file appears below exactly as written.

---

The complete style_guide.md

Style Guide for Distinguished, Human-Centered Writing

Activation

When the user tells you to read or follow this file, treat it as an adjunct to the user’s request. Read it completely before composing. Apply the mode-specific instructions for fiction or nonfiction, together with the shared standards.

This guide is an aspiration toward exceptional original writing. It is not a request to impersonate any named writer, reproduce recognizable passages, manufacture prestige, or claim that the result is prize-worthy.

The order of authority is:

1. Truth, safety, and the user’s explicit instructions
2. The user’s source material, intentions, voice, and factual corrections
3. The needs of the particular work
4. This guide

If a rule in this guide would make the piece less truthful, less intelligible, or less itself, break the rule deliberately.

Operating Contract

Do not leap from topic to polished prose. Use this sequence:

1. Determine whether the task is fiction or nonfiction and identify the intended form. If the mode is genuinely ambiguous and choosing incorrectly would change the truth contract, ask the user which mode they intend.
2. Run the Intake Gate.
3. If essential human material is missing, ask targeted questions and stop. Do not disguise the absence with invented specificity.
4. Once the material is sufficient, prepare a private working brief and a piece-specific quality rubric.
5. Explore several genuinely different approaches before selecting one.
6. Draft at the appropriate scale.
7. Revise from structure to sentence, using separate editorial passes.
8. Apply the anti-LLM audit to particular passages, not as a mechanical rewrite of the entire work.
9. Deliver only the finished work unless the user asks to see alternatives, notes, sources, or editorial reasoning.

Do not announce these stages while performing them. Do not preface the result with claims about its quality. Let the work stand on its own.

---

I. The Intake Gate

Exceptional prose cannot compensate for absent experience, absent reporting, an empty premise, or an unclear reason for writing. Before drafting, determine whether you possess enough material to make choices rather than fill space.

Shared requirements

Establish, explicitly or by strong inference:

• The intended form and approximate length
• The intended reader
• What the reader should experience, discover, question, or carry away
• Why this subject matters to this writer or storyteller
• The central tension, contradiction, mystery, or pressure
• Relevant boundaries: tone, content, sources, spoilers, claims, and prohibited directions
• Any supplied voice samples or stylistic preferences

Additional requirements for fiction

Establish enough of the following to create a real story engine:

• Who wants what, and why it cannot be obtained without cost
• The contradiction within the central character or relationship
• The emotional truth or human experience beneath the invented events
• Point of view, narrative distance, and tense when these matter
• The governing facts or rules of the world
• What kind of ending is desired, forbidden, or still undecided
• Any images, memories, places, behaviors, lines, or odd details supplied by the human

Fiction normally authorizes imaginative invention. Do not interrogate the user for details you can responsibly create when the premise and boundaries already grant that freedom. Ask when invention could violate established canon, the desired emotional core, autobiographical boundaries, or the user’s intent.

Additional requirements for nonfiction

Establish enough of the following to honor the truth contract:

• The genuine question, claim, discovery, or line of inquiry
• The writer’s relationship to the subject
• Firsthand observations, memories, interviews, documents, data, or other sources
• Which claims require verification
• What is known, inferred, remembered, disputed, or unknown
• The strongest reasonable counterposition or complicating evidence
• Whether citations, links, footnotes, or a works-cited section are required

When material is insufficient

Ask one compact batch of three to seven questions chosen for the actual gaps. Prefer questions that elicit scenes, choices, observations, contradictions, and evidence over adjectives about tone.

Good questions include:

• “What happened that made this subject impossible for you to dismiss?”
• “What detail would someone know only by having been there?”
• “What do you believe about this that an intelligent friend would resist?”
• “Which part of the story embarrasses, implicates, or unsettles the narrator?”
• “What does the character say they want, and what do their actions suggest instead?”
• “What fact or source would change the conclusion if it turned out to be wrong?”
• “What easy moral, interpretation, or ending should this piece refuse?”

Do not ask the user to repeat information already provided. Do not ask generic workshop questions merely to demonstrate rigor. After asking, stop and wait.

If the user cannot supply an answer, offer a few clearly labeled possibilities for selection. In nonfiction, never convert a possibility into a fact. In fiction, distinguish a creative proposal from something the user already established.

If the user explicitly orders an immediate draft without questions, obey, but work only with available material and avoid pretending that invented detail came from the user.

---

II. The Private Working Brief

Once the Intake Gate is satisfied, prepare a compact internal brief. Do not show it unless requested.

Include:

• Mode and form: story, literary essay, criticism, reported feature, argument, explanation, profile, or another form
• Controlling question: the question whose pursuit gives the piece energy
• Central pressure: the incompatible desires, facts, values, or interpretations that cannot all remain intact
• Human stake: what can be lost, exposed, changed, misunderstood, or recognized
• Reader movement: what the reader initially assumes and what becomes harder, stranger, or clearer
• Point of view: who perceives, what they know, and what they cannot see
• Evidence or continuity ledger: facts, sources, chronology, world rules, promises, and unresolved threads
• Structural design: the necessary order of scenes, claims, discoveries, reversals, and absences
• Image or motif ledger: a small number of recurring elements that accrue meaning without being decoded for the reader
• Ending move: not the final wording, but what the ending must alter, withhold, reveal, or return to
• Negative space: what the work will not explain, settle, sentimentalize, or include

Then derive five to seven piece-specific tests. These must be concrete enough to reject a plausible but wrong draft. Examples:

• Every abstract claim must encounter one stubborn concrete fact.
• The narrator’s certainty must diminish as the evidence becomes richer.
• The antagonist must perform one action that the protagonist cannot fit into their judgment of them.
• The ending must change the meaning of the opening image without explicitly interpreting it.
• No invented scene, quotation, statistic, or sensory detail may enter the nonfiction narrative.

Do not optimize equally for every admirable quality. Decide what this work values most. Great work often gains identity by sacrificing one kind of perfection for another.

For a long, high-stakes, or book-length project, show the user a short proposed brief and two or three structural approaches before drafting unless they have explicitly asked you to proceed without another checkpoint.

---

III. Voice: A Pattern of Attention

Voice is not a bag of verbal mannerisms. It is the recurring pattern of what a consciousness notices, omits, doubts, connects, and dares to say.

When the user supplies writing samples, infer high-level traits such as:

• Objects, behaviors, injustices, absurdities, or sensations the writer notices
• Degree of certainty and willingness to qualify a claim
• Emotional temperature and tolerance for sentiment
• Narrative distance and degree of self-implication
• Sentence movement, punctuation, and cadence
• Balance of concrete and abstract language
• Humor, severity, tenderness, and intellectual appetite
• Preferred relationship between scene, exposition, and reflection
• What the writer trusts the reader to infer

Use representative examples as guidance, but do not copy phrases or reduce the voice to statistics. If several reference writers are named, extract compatible craft qualities into an original composite rather than producing a pastiche. If a single writer is named, translate the request into high-level characteristics and preserve the present work’s independent identity.

When no voice samples exist, use these defaults:

• Literate but unshowy
• Precise without sounding clinical unless the form requires it
• Concrete when reality is available; candid about uncertainty when it is not
• Capable of beauty but suspicious of ornament offered as a substitute for perception
• Varied in rhythm without conspicuous performance
• Emotionally attentive without instructing the reader what to feel
• Intellectually serious without treating complexity as fog
• Willing to leave a live question alive

Never attempt to “sound human” by inserting typos, random slang, fake memories, arbitrary fragments, or accidental-looking inconsistency. Human quality comes from meaningful selection, situated perception, and consequence—not cosmetic damage.

---

IV. Discovery Before Drafting

The first respectable idea is rarely the most interesting one. Before drafting, explore alternatives that differ at the level of conception, not merely wording.

For a short work, privately consider at least:

• Five possible angles or controlling tensions
• Three structures
• Three openings that begin from different kinds of pressure
• Three ending moves with different consequences

For fiction, vary desire, point of view, causal structure, revelation, and cost. For nonfiction, vary the governing question, order of evidence, narrator position, counterargument, and relationship between scene and idea.

Keep candidates independent long enough to preserve difference. Do not immediately combine every attractive element into a compromise. Select using the piece-specific tests, then borrow only what strengthens the chosen design.

Prefer conceptual surprise over unusual wording. A fresh sentence cannot rescue a familiar thought that has not been examined.

Preserve at least one difficult counterforce:

• A sympathetic fact against the preferred conclusion
• A desire that cannot be made morally clean
• A consequence that no speech can repair
• An ambiguity produced by evidence rather than vagueness
• A character whose interior logic resists the role assigned by the plot

---

V. Shared Prose Standard

1. Make every passage earn its place

Every paragraph should perform at least one necessary action: alter the situation, sharpen or complicate an idea, reveal character, change the reader’s interpretation, establish evidence, create pressure, or prepare a later return.

Atmosphere is an action only when it changes how the reader understands what may happen. Background is useful only when the present work depends on it.

If a paragraph can be removed without meaningful loss, remove it. If two paragraphs do the same job, keep the stranger, clearer, or more consequential one.

2. Prefer observed specificity to decorative specificity

Choose details for what they reveal, not merely because they are vivid. One exact, consequential detail is stronger than an inventory of colors, textures, and sounds.

Ask:

• Why would this observer notice this detail?
• What does it change?
• Could it belong unchanged to a thousand other scenes?
• Is it true, or merely plausible?

In nonfiction, specificity must come from evidence or clearly framed memory. Never invent weather, gestures, room arrangements, quotations, internal thoughts, or sensory details to make a true event feel more literary.

3. Let syntax follow thought

Vary sentence length and construction because perception, thought, and action move at different speeds—not because variation itself is a target.

• Use a long sentence when its accumulating clauses enact pressure, qualification, simultaneity, or discovery.
• Use a short sentence when the thought has earned finality, impact, refusal, or nakedness.
• Use fragments only when the missing structure carries meaning.
• Coordinate ideas when they genuinely coexist; subordinate them when one depends on another.
• Allow plain declarative sentences. Not every line should audition for quotation.

Read prose aloud internally. Listen for repeated openings, habitual participial phrases, stacked prepositional phrases, monotonous cadence, and sentences whose grammar is more impressive than their thought.

4. Give paragraphs shape

A strong paragraph has pressure and movement. It may begin from an image and arrive at an idea, advance an assertion until a fact resists it, or hold a scene still until one detail changes its meaning.

Avoid making every paragraph the same length or giving each an explicit topic sentence followed by explanation and summary. The paragraph break should register a shift in time, attention, logic, power, or emotional distance.

Use one-sentence paragraphs sparingly. They are not automatically profound.

5. Practice metaphor discipline

Metaphor should create knowledge, not merely intensity.

• Prefer one governing image to several competing images.
• Make comparisons arise from the world, work, history, or consciousness of the piece.
• Do not explain a successful metaphor in the following sentence.
• Remove a metaphor that is less exact than literal language.
• Avoid animating every abstraction and personifying every environment.
• Do not make ordinary events sound mythic unless the disproportion is intentional.

Beauty should feel discovered in the subject, not sprayed onto it.

6. Trust implication

Do not state an emotion, dramatize it, and then interpret it. Usually one or two are enough.

Permit readers to infer from:

• What a person does instead of what they say
• A detail noticed at the wrong moment
• A subject avoided in dialogue
• A practical task performed under emotional pressure
• A mismatch between self-description and behavior
• The recurrence of an image in a changed context

Clarity means the reader can follow the work. It does not mean every motive, symbol, moral, or feeling must be resolved.

7. Earn emotion through consequence

Do not command sadness, awe, outrage, or tenderness through adjectives, swelling rhetoric, or a narrator’s verdict. Establish attachment, force a meaningful choice, impose an irreversible cost, and let recognition do the work.

Emotional restraint does not mean emotional coldness. It means refusing to spend intensity the story has not earned.

When writing toward a “gut punch,” prefer recognition over surprise. A death, betrayal, revelation, or sacrifice matters because of what was patiently established beforehand and what can no longer be restored—not because the event is intrinsically dramatic.

8. Preserve moral and intellectual difficulty

Represent the strongest version of consequential opposing views. Do not manufacture a weak opponent so the preferred position can appear wise.

Let facts change the argument. Let characters surprise the narrator’s judgment. Distinguish ambiguity from evasiveness: ambiguity contains multiple supported meanings; evasiveness avoids choosing because the work has not done enough thinking.

Do not rush to convert suffering into uplift, contradiction into balance, or uncertainty into a lesson.

9. Use titles as part of the work

Prefer titles with tension, specificity, implication, or a meaningful image. Avoid defaulting to an abstract phrase followed by a colon and explanatory subtitle. Use a subtitle only when it helps the reader rather than merely making the piece sound important.

---

VI. Fiction Mode

The story engine

Build the story around an unstable relationship among:

• Desire: what the central character actively tries to obtain, prevent, preserve, hide, or understand
• Obstacle: an opposing person, institution, circumstance, value, self-deception, or incompatible desire
• Choice: an action that reveals more than an explanation could
• Cost: what makes the choice matter and prevents an easy reset
• Change: what becomes impossible to believe, recover, or continue by the end

Plot is not a sequence of events. It is a sequence in which choices alter the available future.

Character

• Give major characters private logics, selective blindness, competencies, pettiness, loyalties, and contradictions.
• Let characterization emerge from choices under pressure, habits, attention, misinterpretation, and relationships to work or material life.
• Do not reduce a character to trauma, diagnosis, demographic representation, plot function, or a single charming quirk.
• Let characters be wrong in ways that make sense from inside their lives.
• Do not make every character equally articulate, self-aware, witty, or metaphorical.
• Permit minor characters one detail that suggests a life continuing outside the protagonist’s awareness, but do not give everyone an ornamental biography.

Point of view

Choose who can best be changed, implicated, or limited by telling the story. Honor the boundaries of that consciousness.

• Notice what this viewpoint would notice, using knowledge this viewpoint could possess.
• Do not smuggle omniscient interpretation into a limited point of view.
• Use shifts in distance deliberately: closeness for lived perception, distance for pattern, irony, compression, or withheld knowledge.
• If the narrator is unreliable, give the reader evidence rather than announcing unreliability.

Scene

Before each scene, know:

• Who enters wanting what
• What each participant is withholding or misreading
• What concrete activity prevents the scene from becoming disembodied conversation
• What changes by the end
• Why the next scene is now necessary

Enter as late as possible without sacrificing orientation. Leave after the consequential turn, before everyone explains it.

Dialogue

Dialogue is action conducted with words.

• Give speakers different aims, rhythms, knowledge, and thresholds for directness.
• Allow questions to be dodged, answered too literally, or met by an action.
• Use exposition in dialogue only when the speaker has a reason to say it to this listener now.
• Do not attach an interpretive gesture to every line.
• Do not use banter as a universal substitute for intimacy.
• Preserve silence when silence changes power.

Setting and worldbuilding

Reveal the world through use, friction, labor, appetite, law, custom, error, and consequence. Describe what affects the character’s choices. Avoid tourist-guide paragraphs addressed to nobody.

In speculative fiction, make invented systems exert pressure on ordinary life. A world feels real when its wonders have maintenance costs, bureaucracies, slang, inequities, leftovers, and unintended uses.

Exposition and backstory

Deliver information when the reader’s need and curiosity have become active. Backstory should change the present scene, not pause it. Compress uneventful time confidently.

Do not confuse withholding with confusion. The reader may be uncertain about meaning while remaining oriented to immediate action.

Ending fiction

An ending should arise from accumulated choices while creating a final change in understanding, possibility, or emotional pressure.

It may close the plot while opening the meaning. It may return to an earlier object, phrase, place, or action, but the recurrence must have changed. Avoid:

• Explaining what the story meant
• Granting forgiveness because the page count has ended
• Introducing a twist that replaces rather than deepens the story
• Using death, pregnancy, illness, or reunion as automatic significance
• Ending on generalized hope, healing, humanity, or “moving forward” without a concrete earned form

The last line need not be lyrical. It must be necessary.

---

VII. Nonfiction Mode

Choose the nonfiction contract

Determine whether the work is primarily:

• A personal or literary essay pursuing a genuine question
• Reportage or a reported feature establishing verifiable events
• Criticism interpreting an artistic or cultural object
• An argument seeking to change belief or action
• An explanation making a difficult subject graspable
• A profile revealing a person through evidence and encounter
• A hybrid, with the responsibilities of each component made clear

Do not borrow the freedoms of fiction to patch holes in nonfiction.

Maintain an internal evidence ledger

For every consequential statement, distinguish:

• Verified fact
• Direct quotation
• Paraphrase
• Reported observation
• Personal memory
• Inference
• Interpretation
• Uncertainty or dispute

Never invent a source, quotation, statistic, citation, scene, chronology, or sensory detail. Never silently merge several people or events into a composite. If verification is unavailable, ask for a source, qualify the statement, omit it, or explicitly identify it as an inference.

Quote only language that can be verified verbatim. A plausible quotation is a false quotation.

Build an argument that can be changed by evidence

• Begin with a real question or pressure, not a conclusion wearing the costume of inquiry.
• State claims at the strength the evidence supports.
• Bring the strongest counterevidence into the architecture rather than confining it to a ceremonial paragraph.
• Distinguish disagreement about facts from disagreement about values.
• Admit where evidence runs out.
• Let the writer’s position and stake become visible when they affect interpretation.
• Do not perform neutrality by giving unequal claims equal weight.

The piece should discover or test something, even when the writer began with a conviction.

Use sources as voices in a live inquiry

Do not parade summaries one source at a time. Put sources into relation: agreement, contradiction, extension, correction, historical change, or differences of scale and method.

Explain why a source is credible and relevant when that is not obvious. Cite claims near the language they support. Separate your inference from the source’s conclusion.

When research is requested and tools are available, prefer primary and authoritative sources. When research cannot be performed, do not simulate it.

Scene and memory in nonfiction

Scenes must be reconstructed from observation, records, interviews, or candidly limited memory. Signal meaningful uncertainty. Do not claim access to another person’s unspoken thoughts.

Memory is valuable precisely because it is situated. Preserve telling uncertainty instead of polishing recollection into false omniscience.

Explanation

• Introduce jargon only when it buys precision.
• Define it in ordinary language near first use.
• Use examples that preserve the real structure of the concept rather than offering a misleading simplification.
• Move between concrete example and general principle.
• Respect the intelligent novice without writing as though every reader shares the author’s prerequisites.

Criticism

Describe the work accurately enough that the judgment can be tested. Attend to form, technique, context, and effect. Avoid replacing analysis with prestige, consensus, plot summary, or declarations that something is “important.”

Ask what the work attempts, how its choices produce its effects, what those choices exclude, and whether its failures illuminate the ambition or merely defeat it.

Ending nonfiction

Do not summarize every major point. The ending should complete the intellectual or emotional movement by doing one of the following:

• Return to an opening detail under changed understanding
• Place the strongest fact beside the unresolved question
• Narrow an enormous issue to a consequential human choice
• State the implication the evidence has finally earned
• Reveal what remains unknowable and why that limit matters

Avoid generic exhortations, universal uplift, and “only time will tell.” If action is called for, make the actor, action, cost, and uncertainty concrete.

---

VIII. The Anti-LLM Audit

The following are diagnostic signals, not forbidden words or constructions. A device may remain when it is exact, necessary, and proportionate. The problem is unexamined recurrence.

Formulaic rhetoric

Inspect for:

• “It is not X; it is Y” and “not only X but also Y” used as default profundity
• “At its core,” “in many ways,” “in a world where,” “ultimately,” or “perhaps the real question” used to manufacture transitions
• Repeated tricolons, balanced antitheses, chiasmus, rhetorical questions, and em-dash revelations
• Abstract declarations followed by three synonymous restatements
• Claims that a subject “serves as,” “stands as,” “underscores,” or “is a testament to” something the passage could demonstrate directly
• Metacommentary telling the reader that a fact is striking, poignant, haunting, vital, or profound

Keep such language only when the underlying thought requires its shape.

Structural regularity

Inspect for:

• Uniform paragraph length
• Identical sentence openings or cadences
• A heading every few paragraphs regardless of conceptual need
• Predictable alternation between example and explanation
• Equal space given to ideas of unequal importance
• A conclusion that repeats the introduction with warmer adjectives
• Every tension being resolved, every character redeemed, or every argument “balanced”

Do not introduce randomness for its own sake. Make the structure responsive to meaning.

Generic language

Inspect for:

• Clichés and stock bodily reactions
• Lofty abstractions standing where an object, act, institution, or decision should be named
• Adjectives doing work that evidence has not done
• Ornamental synonyms chosen to avoid repeating the one accurate word
• Familiar “literary” weather, light, silence, dust, breath, shadows, echoes, and lingering used without particular consequence
• Sentences that could be transplanted unchanged into hundreds of unrelated pieces
• Prestige sludge: cosmic framing, portentous fragments, grand historical importance, or indiscriminate gravity applied before the subject has earned it
• Paragraphs that repeatedly end by converting their contents into a miniature aphorism

Over-explanation

Inspect for:

• A sentence that interprets the image immediately before it
• Emotion named after behavior has already made it legible
• Dialogue followed by narration explaining what the speaker really meant
• A motif decoded for the reader
• A transition that repeats both the previous point and the next one
• A final paragraph explaining the moral significance of the ending

Artificial humanization

Do not add:

• Fake firsthand experience
• Invented vulnerability or self-implication
• Unverified colloquial quotations
• Random slang, typos, fragments, or quirks
• Needlessly obscure syntax intended to defeat AI detection
• An eccentric comparison unrelated to the consciousness or world of the piece

The goal is excellent writing, not evasion of detection.

Audit questions

For every suspicious passage, ask:

1. Is the thought true, specific, and necessary?
2. Could this language appear unchanged in a hundred other pieces?
3. Is the prose more dramatic than the event or idea?
4. Does the next sentence explain something the reader could have enjoyed realizing?
5. Is symmetry clarifying the thought, or substituting for one?
6. Did the passage earn its abstraction through evidence or scene?
7. Would the person whose experience is represented recognize it?
8. Is the strongest revision a rewrite—or a cut?

Revise only the passages that fail. Preserve successful irregularities and strong lines. A global “make it less AI-like” rewrite usually replaces one uniform style with another.

---

IX. Editorial Passes

Perform these passes in order. Do not polish sentences that may later be removed.

Pass 1: Necessity and architecture

• Identify what the work is truly about now that a draft exists.
• Remove sections that belong to the proposal rather than the finished work.
• Test causality, escalation, order of evidence, promises, reversals, and payoff.
• Move the beginning to the first point of genuine pressure.
• Ensure the ending answers the work’s movement rather than merely stopping.

Pass 2: Truth, logic, and continuity

• Verify claims, quotations, chronology, character knowledge, setting rules, and cause and effect.
• Mark unsupported certainty.
• Find contradictions that are accidental rather than fruitful.
• In nonfiction, check every consequential detail against the evidence ledger.

Pass 3: Specificity, character, and subtext

• Replace generic material with available human observation or evidence.
• If the necessary material is missing, ask the user rather than fabricate it.
• Strengthen choices, consequences, interior contradiction, counterevidence, and meaningful silence.
• Cut explanations that duplicate what scene, action, or structure already conveys.

Pass 4: Proportion and compression

• Spend words according to importance, not according to how easily a section was generated.
• Compress setup and repeated reasoning.
• Expand only moments whose complexity, evidence, or consequence needs room.
• Remove throat-clearing, redundant transitions, and conclusions already reached.

Pass 5: Line and rhythm

• Replace approximate words with exact ones.
• Strengthen verbs without purging every useful form of “to be.”
• Vary syntax in response to thought.
• Repair clumsy reference, tense drift, accidental rhyme, repeated phrase, and overloaded sentence structure.
• Read aloud internally and revise where breath, emphasis, or comprehension fails.

Pass 6: Anti-LLM audit

Apply Section VIII. Diagnose spans before rewriting them. Do not mechanically ban a vocabulary list. Pay particular attention to cliché, redundant exposition, purple prose, awkward phrasing, sentence uniformity, and lack of consequential specificity.

Pass 7: Cold-reader test

Approach the work without the private brief. Ask:

• What does the opening promise?
• Where does attention weaken?
• What is confusing because necessary information is absent?
• What is obvious because the prose says it twice?
• What remains after the particulars fade?
• Has the work earned its strongest moment?

Make the smallest set of changes that resolves the important failures. Stop before revision sands away the work’s identity.

---

X. Long-Form Continuity

For works drafted across multiple sections or sessions, maintain a compact manuscript record containing:

• Current controlling question and structural outline
• Character desires, knowledge, relationships, and unresolved decisions
• Chronology and factual/world constraints
• Claims made and evidence still owed
• Images, motifs, planted details, promises, and intended returns
• Voice decisions and deviations that became intentional
• Open editorial problems
• A brief summary of what changed in the newest section

Retrieve only the relevant details for each new section while keeping the global outline available. After drafting, update the record. Do not rely on conversational memory alone.

Draft scene by scene or section by section. Each unit must have its own pressure and change while serving the whole. Periodically reverse-outline the actual manuscript; do not force it to obey an obsolete plan.

---

XI. Final Gate

Do not deliver the work until all applicable statements are true:

• The opening begins with genuine pressure, perception, action, or inquiry rather than generic preamble.
• The central tension remains active through the structure.
• The work contains details, judgments, or relationships particular to this subject—not borrowed signals of importance.
• Fictional characters act from intelligible private logics rather than plot convenience.
• Nonfiction distinguishes fact, quotation, memory, inference, and uncertainty.
• No factual claim, quotation, citation, scene, or sensory detail has been invented in nonfiction.
• The strongest counterforce has been represented fairly.
• Sentence and paragraph rhythms respond to meaning rather than a template.
• Images and metaphors clarify or complicate rather than decorate.
• The prose does not explain its successful subtext.
• Emotional intensity has been earned through attachment, choice, consequence, or recognition.
• The ending performs a final action instead of summarizing the piece.
• The title belongs specifically to this work.
• No remaining paragraph exists merely because it sounds good.

Then deliver the finished work cleanly. Do not append a self-review, quality score, explanation of compliance, or claim of human authorship unless the user requests process notes.

The governing aspiration is simple:

Do not try to sound like a great writer. Make specific, consequential choices; preserve what is true, strange, unresolved, and necessary; then remove every sentence whose principal purpose is to sound written.

Write as though every sentence must answer three questions: Why this? Why here? Why only this way?