../

Writing

Writing well with an AI as collaborator and editor rather than ghostwriter, across nonfiction (essays, explainers, reports, memos), persuasive, technical and fiction writing, plus the four levels of editing and how to ask for each. The emphasis throughout is on keeping your own voice: the prompt chain takes your draft to final through critique, and you hold the pen at every pass. It applies the prompt skeleton and method in prompting fundamentals; the craft itself is in the writing sheets, especially clear writing and story structure. Producing marketing content at volume across channels is a different job, covered in content generation.

How experts think

Writers and editors share a few convictions that change how you should use a model:

  • Writing is thinking. If you cannot state the point in one sentence, you are still thinking. A model can help you think (by interviewing you, arguing back, asking what you mean), but if it writes the first draft it also does the thinking, and the piece becomes its, not yours.
  • Reader first. Every choice serves a specific reader: what they know, what they need, what they will do next.
  • Revision is most of the work, done in passes from large to small: structure, then paragraphs, then sentences, then words, then typos. Polishing sentences in a section you will cut is wasted effort.
  • Editors diagnose; authors fix. A good editor says what isn't working and why, and leaves the wording to the author. That is how the author's voice survives an edit.
  • Voice is a set of choices: which words, what rhythm, what you notice, what you leave out, where the jokes go. Models drift toward the average word and the average rhythm, so unguarded edits flatten voice.

Give the model one named job per session rather than "help me write", and put the job in the prompt's <task>:

JobUse it whenRisk
Interviewerblank page; before outlininglow: the ideas stay yours
Sparring partnertesting an argument or premiseit may raise objections no real reader would
Simulated readerchecking clarity for a named audienceit is not your real reader; confirm with people
Developmental editstructure, argument, plotpushes toward formula if you let it
Line editsentence-level style and flowflattens voice if allowed to rewrite freely
Copyeditgrammar, consistency, style rulesintroduces errors in names, quotes and technical terms
Proofreadthe last pass for typosmisses some errors; "fixes" things that were right
Ghostwritingyou truly don't mind whose voice it isthe piece sounds like everyone else's

Vocabulary

Nonfiction and persuasion

TermWhat it meansSay it like
Thesis (controlling idea)the one claim a piece exists to make"State my thesis in one sentence. If you can't, tell me why."
Target readerthe specific person the piece is written for"Read this as a CFO with ten minutes and no context"
BLUF"bottom line up front": the conclusion or request first"Rewrite the opening as BLUF: the ask, then the reason"
Lede / nut grafan article's opening / the paragraph that says why the story matters"Is the nut graf clear by paragraph three?"
Topic sentencea paragraph's first sentence, stating its one idea"List each paragraph's topic sentence; flag paragraphs without one"
Signpostingphrases that tell the reader where the argument is going"Add signposts where a reader could get lost; no more than needed"
Through-linethe thread connecting every section to the main point"Which sections break the through-line?"
Registerthe formality level of the language"Keep the register conversational; flag anything stiff"
Dictionword choice"Flag diction that's vaguer or fancier than it needs to be"
Cadencethe rhythm of sentences: length, stress, pauses"Point out runs of same-length sentences; don't fix them"
Hedgea qualifier that weakens a claim ("perhaps", "somewhat")"List my hedges; mark which are honest uncertainty and which are nerves"
Claim, evidence, warrantfrom Stephen Toulmin's model of argument (he called evidence "data", and later "grounds"): the assertion, the support, and the reasoning that links them"For each claim, name the evidence and the unstated warrant"
Steelmanthe strongest version of the opposing view"Steelman the case against my argument in 150 words"
Concession and rebuttalgranting a valid point, then answering it"Where should I concede, and where rebut?"
Pyramid principle / SCQABarbara Minto's structure: answer first, then supporting groups; intros as situation, complication, question, answer"Restructure as a pyramid: answer, three reasons, evidence"
Diátaxisa framework that splits documentation into tutorials, how-to guides, reference and explanation"Which Diátaxis type is this page trying to be? It mixes two."

Fiction

TermWhat it meansSay it like
Premise / loglinethe story's core situation / a one-sentence pitch of it"Give me five loglines for this premise; keep my protagonist"
Inciting incidentthe event that disrupts the status quo and starts the story"Where does my inciting incident happen, and is it too late?"
Stakeswhat the character stands to lose"Are the stakes clear and personal by the end of chapter 2?"
Want vs needthe external goal a character pursues vs the inner change they require"Name each lead's want and need from the draft alone"
Archow a character changes (or deliberately doesn't)"Trace her arc scene by scene; where does it stall?"
Scene and sequelDwight Swain's pairing: goal, conflict, disaster; then reaction, dilemma, decision"Label each scene's goal, conflict and disaster"
Beatthe smallest unit of action and reaction in a scene"Break this scene into beats; mark the turn"
POV (point of view)whose perspective narrates: first, second, third limited or omniscient"Check this chapter for POV slips in close third"
Psychic distanceJohn Gardner's term for how close the narration sits to a character's mind"Mark where psychic distance jumps and whether it's deliberate"
Free indirect stylethird-person narration colored by a character's words and attitudes"Which sentences are free indirect? Keep them; they're the voice"
Head-hoppingswitching minds mid-scene in limited third, usually by accident"Flag any head-hopping"
Tensepast, present or future narration"Flag every tense slip; don't change anything else"
Narrative voicethe personality of the telling: diction, rhythm, attitude"Describe the narrator's voice; then flag lines that break it"
Show vs tellrendering in action and detail vs stating conclusions; both have uses"Mark telling at emotional peaks only; summary elsewhere is fine"
Subtextmeaning beneath what characters say"Where does dialogue say the subtext out loud?"
Filter words"she saw", "he felt", "she noticed": they put a layer between reader and experience"List filter words in this scene with the sentence"
Exposition / info-dumpbackground information / too much of it at once"Where can this exposition be cut or moved into action?"
"As you know, Bob"dialogue where characters tell each other what both already know, for the reader's benefit"Flag any 'as you know, Bob' dialogue"
Setup and payoff (Chekhov's gun)anything prominent early must matter later, and payoffs need setups"List setups without payoffs and payoffs without setups"
Unreliable narratora narrator whose account the reader learns to distrust"Is the unreliability detectable on a second read?"
psychic distance: how close the narration sits to her mind In the winter of 1921, a woman left the station. far: what a camera would see; no name, no thoughts Mara Voss had never much liked trains. named, summed up by a narrator outside her Mara hated trains. her feeling, reported God, how she hated these trains. her words inside the narration (free indirect style) Trains. The smell, the shudder, the endless — inside her mind: stream of consciousness Move between steps on purpose: an unplanned jump of several steps mid-scene reads as a slip. after John Gardner's five-step example
Psychic distance, after John Gardner's five-step example in The Art of Fiction: pick a step for each scene and move between steps on purpose.

Editing

TermWhat it meansSay it like
Developmental editbig-picture edit of content, structure and argument or plot (also: substantive, structural)"Developmental edit only: comments, no line changes"
Editorial lettera developmental editor's written report on the whole draft"Write an editorial letter: strengths, top three problems, suggested directions"
Line editsentence- and paragraph-level edit for clarity, flow and style"Line edit this section; keep my rhythm and vocabulary"
Copyeditedit for grammar, spelling, punctuation, usage and consistency"Copyedit to Chicago style and give me a style sheet"
Proofreadfinal check of the finished text for typos and formatting errors"Proofread only: list errors with locations"
Style sheetthe record of spelling, capitalization and usage decisions for one manuscript"Start a style sheet: names, terms, numbers, hyphenation"
Queryan editor's question to the author instead of a change"Where you're unsure what I meant, query; don't guess"
Tracked changes (redline)edits shown as visible insertions and deletions the author accepts or rejects"Show every change as tracked: old new"
Reverse outlinean outline made after drafting, one line per paragraph or scene"Reverse-outline the draft, one line per paragraph"
Beta readera test reader who reports their reactions, not edits"React as a first-time reader: where were you bored, confused, moved?"
Stet"let it stand": restore the original after a suggested change"I've marked stet on changes I rejected; don't re-suggest them"

Keep your own voice

Set the editor rules once

Paste these at the start of a session, or save them in your assistant's project or custom instructions.

EDITOR RULES
- Edit; don't ghostwrite.
- Diagnose before you prescribe: say what isn't working and
  why, quoting the passage.
- Default to comments. Rewrite only when I ask, and only the
  passage I name.
- When you suggest wording, offer it as an option, keep my
  vocabulary and sentence rhythm, and change as little as
  possible.
- Never change meaning, facts, quotes, names, technical terms
  or dialect in dialogue.
- Never add claims, examples or anecdotes I didn't give you.
- Show every change so I can accept or reject it.
- Label matters of taste as taste, and errors as errors.
- Tell me what is working, briefly, so I don't cut it.

Give samples, not adjectives

"Friendly but authoritative" means nothing to a model. Paste a few pages of your own finished writing and ask it to describe your voice in specifics first: average sentence length, how paragraphs end, favorite constructions, what you never do. Check the description, correct it, and reuse it. Use the description to guide edits; be wary of "write like me" imitations, which copy surface tics and miss judgment.

Ask for critique instead of rewrites

Output formatGood forAsk for it like
Comment tabledevelopmental and line feedback"Table: location, problem, why it matters to the reader, direction (no wording)"
Ranked problemsgetting past politeness"The three biggest problems, most serious first"
Questionswhen you might be unclear"Ask me five questions a confused reader would ask"
Optionsstuck sentences"Three versions of this sentence, each changing as little as possible"
REWRITE: "make it better" CRITIQUE: the model diagnoses, you fix your draft the model rewrites every sentence clean text, no record your draft model: quote, problem, why you revise in your own words meaning check: hedges, numbers its voice; you can't see what changed next pass, next level down your voice survives, and every change is one you chose
Editors diagnose; authors fix. The critique loop keeps the pen in your hand.

Get tracked-change style output

MethodHow
Change table"Return a table: original, suggested, reason. Don't return the full text."
Markdown marks"Show deletions as struck and insertions in bold, inline."
CriticMarkupa plain-text syntax for edits: {--deleted--}, {++added++}, {~~old~>new~~}, {>>comment<<}
Compare documentsask for the clean revised text, then compare it with your original in your word processor or a diff tool to see every change
Suggestion modepaste accepted changes into a shared document in suggestion mode so collaborators can review them

Preserve meaning

Edits fail quietly by changing what you said. Watch for: hedges removed ("may" becomes "will"), qualifiers dropped, numbers rounded, a nuance "simplified" away, a claim strengthened for punch. After any edit, run the meaning check in Judge the output.

Editing levels

Definitions follow the Editorial Freelancers Association (EFA). Do the levels in order; each assumes the one before is settled.

LevelAlso calledFixesWhenAsk for it likeIt does not
Developmental editsubstantive, structural, content editcontent, organization, argument or plot, genre fit; delivered as an editorial letter or manuscript evaluationafter a complete draft"Developmental edit: big-picture comments and an editorial letter; no line changes"fix sentences
Line editstylistic editlanguage at the sentence and paragraph level: clarity, flow, rhythm, word choice, voiceonce structure is settled"Line edit section 2 as tracked changes; keep my voice"restructure
Copyedit—spelling, grammar, usage, punctuation, consistency, cross-references; produces a style sheetwhen content is final"Copyedit to [style guide]; queries for anything ambiguous; style sheet at the end"change style
Proofread—typos and formatting errors in the final, laid-out textlast"Proofread only: list each error with its location; change nothing else"edit
In order to find out why, we emailed 12 clients; ten replied the the same day. To we emailed 12 clients; ten replied the the same day. 10 style sheet: numerals for 10+ the the doubled word DEVELOPMENTAL structure, argument, plot: move, cut, add LINE EDIT clarity, flow, rhythm, voice COPYEDIT grammar, usage, consistency PROOFREAD typos left in the final layout in this order: each level assumes the one before is settled
From the whole draft down to a single word. Polishing sentences in a section you will later cut is wasted work.

Related services: a manuscript critique (a report without line edits); a beta read (reader reactions); a fact-check (verifying claims against sources); and sensitivity or authenticity reads (how a group is portrayed), which need readers with lived experience; a model is no substitute.

Forms

Nonfiction

FormIt mustGive the modelAsk for
Essayexplore or argue one idea, with a personal angleyour thesis, your examples, the question you're working outan interview; counterarguments; structure options; a reverse outline
Explainertake a newcomer from zero to understandingthe concept, the reader's starting point, where people get confusedan outline from simple to complex; analogy options to vet; questions a newcomer would ask next
Reportpresent findings and what they meandata, methods, the decision it informsan executive summary; a findings-first structure; chart suggestions
Memo or emailget a decision or actionthe ask, the deadline, the contexta BLUF version; subject lines; a cut to one screen

The rules of clear prose (one point, reader first, strong verbs, given-before-new) are in clear writing.

Persuasive

Persuasion is argument plus reader psychology. Use the model to map the argument (claims, evidence, warrants), to steelman the other side, and to list objections in the order a skeptic would raise them; keep the evidence and the conclusion yours. Structures such as the pyramid principle, Amazon-style narrative memos and pitch outlines are in persuasive writing.

Technical

Ask the model to classify each page by Diátaxis type (tutorial, how-to guide, reference or explanation) and to flag mixed pages; ask for procedures as numbered steps with one action each and an expected result. Run every command and check every API name yourself: models invent plausible flags and parameters. Formats (READMEs, design docs, ADRs, runbooks) are in technical writing.

Fiction

Models are good at naming and diagnosing craft problems and weak at original, surprising prose. Left alone they resolve conflict too quickly, have characters explain their feelings, reach for stock phrases ("a breath she didn't know she was holding") and end scenes on a moral.

ElementWhat to ask the modelKeep for yourself
Structuremap your outline to three acts, Save the Cat! beats or try/fail cycles; find sagging middles and missing turnsthe events and the ending
Charactername each character's want, need and arc from the draft alone; flag passive protagonistswho they are
Scenelabel goal, conflict and disaster for each scene; flag scenes where nothing changeswhat happens
POV and distanceflag head-hopping and unplanned jumps in psychic distancethe choice of POV
Tenseflag slipsthe choice of tense
Show vs tellflag telling at emotional peaks, with the sentencehow to show it
Dialogueflag subtext said aloud, "as you know, Bob", identical voicesthe lines
Pacinga scene-by-scene map of scene vs summary and tension levelthe rhythm
Brainstorming20 options for a turn, then discard the first five, which are the most predictablethe pick
inciting incident plot point 1 midpoint plot point 2 climax resolution ACT 1 setup ACT 2 confrontation ACT 3 resolution tension try, fail, raise the stakes sagging middle: nothing turns act lengths of about 1/4, 1/2, 1/4 are typical, not rules
Three-act beats on a tension curve. Ask the model to map your outline onto them and point out where the middle sags.

Structure models are in story structure, people in character, and the scene and sentence level in scenes and prose.

Brief the model

The Slot column says where each input goes in the prompt skeleton.

InputSlotWhy it mattersExample
The draft<materials>long material goes first, before the requestthe full text, or one chapter with a running synopsis
Voice profile<materials>a checkable description of how you writethe output of the voice profile prompt below
Voice samples<examples>the only reliable way to convey your voice"Two finished pieces of mine, about 2,000 words"
Form and genre<context>conventions differ: a memo is not an essay; a thriller is not literary fiction"Personal essay for a general-interest magazine"
Reader<context>sets the bar for clarity and what can go unsaid"Smart non-specialists who have never heard of the topic"
Purpose<context>what the reader should believe or do afterward"Approve a hire by Friday"
Draft stage<context>stops line edits on a draft that needs restructuring"First complete draft; structure not settled"
Known worries<context>focuses attention"I think the middle drags and the ending is rushed"
The job for this session<task>thinking, structuring, critiquing and polishing need different behavior"Developmental feedback only today"
Level of edit<task>"edit this" gets an unhelpful blend of all four levels"Line edit, not a copyedit"
Style rules<requirements>consistency and house style"Chicago Manual of Style; American spelling; serial comma"
What to preserve<requirements>prevents flattening"Keep the jokes, the fragments and the dialect in dialogue"
What not to touch<requirements>protects settled or sensitive parts"Don't touch the quotes or the opening paragraph"
How it will be judged<quality_bar>the model checks its notes before you read themthe checks in Judge the output
Output format<output_format>lets you review fast and stay in control"A table: location, problem, why, suggestion"
How to work<process>one level per pass; questions before guesses"Ask first if my purpose is unclear; no full rewrite"

What to ask for

StageNamed deliverables
Before draftingan interview to surface what you think; one-sentence thesis options; a reader profile; outline options; a logline and beat sheet (fiction); research questions to verify yourself
Drafting (you write)answers to "what would the reader ask next?"; options for one stuck sentence; a list of what the next scene must accomplish
Revisinga reverse outline; developmental notes or an editorial letter; an argument map; a scene-by-scene analysis; a pacing map
Polishinga line edit as tracked changes; a copyedit with a style sheet and queries; a proofreading error list; title options
Checkinga fact-check list; a meaning-preservation check; a consistency check (names, dates, timeline, terms)

The prompt chain

The chain takes a piece of your own writing from draft to final through critique: an editor's intake interview, revision options, the edit, then a review of your revision. You write every sentence; the model diagnoses. Run stages 1 to 3 in one chat, and repeat 3 and 4 one edit level down each time (developmental, then line, then copyedit). Run stage 4 in a fresh chat, because a model reviewing its own suggestions shares its blind spots.

1. Interview

When you have a complete draft: the model asks what an editor asks before taking on a manuscript.

<materials>
[paste your draft]
</materials>
 
<context>
This is my own writing. I want to take it to final myself,
with an editor's critique at each pass. I'm not an editor.
Form: [form: essay | op-ed | report | short story | novel chapter]
</context>
 
<task>
Interview me the way an editor does before taking on a
manuscript. Don't edit, rewrite or assess the draft yet.
</task>
 
<requirements>
- At most 10 questions, grouped: reader and purpose; draft
  stage and worries; voice and what to preserve; style rules
  and limits.
- For each: why the answer changes the edit, and 2-3 typical
  answers I can pick from.
- Use the edit levels (developmental, line, copyedit,
  proofread), each defined in one line the first time.
</requirements>
 
<output_format>
Numbered questions under the four group headings, each
followed by "Why:" and "Typical answers:".
</output_format>
 
<process>
Ask all the questions, then stop. When I answer, restate the
brief as bullets, including which edit level the draft needs
first and why.
</process>

2. Options

When the brief is agreed: most drafts can be revised in more than one direction, so choose the direction before anyone touches a sentence.

<materials>
[paste the brief from stage 1]
[paste your draft]
</materials>
 
<task>
Propose 2-4 materially different revision plans for this
draft, at the developmental level, and recommend one.
</task>
 
<requirements>
- First, say in two sentences what you think the piece is
  about and what it wants the reader to do or feel. If that
  doesn't match my purpose, that's the first finding.
- Plans differ in structure or argument, not wording: e.g.
  restructure around the strongest scene, narrow to one
  claim, open differently, or cut a section.
- For each: what the piece becomes, what it gains, what it
  costs (including what of mine gets cut), effort, and the
  risk to my voice.
- Base every plan on what's in the draft; add no ideas,
  anecdotes or facts.
</requirements>
 
<output_format>
The two sentences, then a table: Plan | The piece becomes |
Gains | Costs | Effort | Voice risk. Then the recommendation
in 3 sentences.
</output_format>
 
<process>
Stop after the recommendation and wait for my choice. Don't
edit anything.
</process>

3. Produce

When you've chosen a plan: the edit, at one level, as notes and tracked suggestions, never a clean rewrite. You revise in your own words.

<materials>
Draft: [paste]
Brief and chosen plan: [paste from stages 1-2]
Voice profile: [paste, from the voice profile prompt]
</materials>
 
<context>
Form: [form, e.g. a 1,800-word personal essay]
Reader: [reader, e.g. smart non-specialists, magazine readers]
Purpose: [purpose, e.g. make readers rethink duty to a parent]
Stage: [stage: first draft | structure settled | content final]
</context>
 
<examples>
[paste 1-2 passages of mine I consider finished, with why]
[paste 1 edit I rejected, with why, e.g. it cut a fragment]
</examples>
 
<task>
Edit the draft to carry out the chosen plan at the
[level: developmental | line | copyedit] level, so that I can
revise it in my own words.
</task>
 
<requirements>
- Diagnose before you prescribe: quote the passage and say
  why it matters to the reader.
- Developmental: an editorial letter and a reverse outline
  with keep, cut, move or expand per paragraph. Line: tracked
  suggestions only where the plan or clarity needs them,
  each as small as possible. Copyedit: to
  [style: Chicago | AP | our house style], with queries and a
  style sheet.
- For narrative: note show-don't-tell at emotional peaks and
  unplanned shifts in POV distance, where they serve the plan.
- Keep my vocabulary, rhythm, fragments and jokes; label
  taste as taste and errors as errors.
- Never change meaning, facts, quotes, names or dialect in
  dialogue; never add claims, examples or anecdotes.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
- Don't return the full rewritten text.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every note stays at the requested level.
- Every note quotes the passage and names the effect on the
  reader.
- Every suggestion keeps meaning, hedges and numbers exactly.
- Read aloud, every suggested phrase fits the voice profile.
- The notes are ranked, biggest problem first.
</quality_bar>
 
<output_format>
1. Three things that are working, so I don't cut them.
2. The notes: editorial letter and reverse outline, or a
table: # | original | suggested | reason | taste or error.
3. Queries where my meaning is unclear, as [AQ: ...].
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made. I'll revise, then run the review.
</process>

4. Review

When you've revised: open a fresh chat and paste the brief, the original and your revision. This is an editor's critique of your work: it scores, quotes and ranks, and never rewrites or supplies wording.

<materials>
Brief: [paste]
Original draft: [paste]
My revised draft: [paste]
</materials>
 
<context>
This is a fresh chat. I revised the draft myself after an
edit. I want an editor's verdict before the next pass.
</context>
 
<task>
Critique my revised draft against the rubric below and rank
what's left to fix. Don't rewrite anything or suggest
replacement wording.
</task>
 
<requirements>
- Score 1-5, with one quoted line of evidence each: purpose
  (does it do what the brief says, for this reader);
  structure; clarity; voice (still mine, compared with the
  original); meaning preserved (every claim, hedge and
  number); nothing invented.
- Fiction or narrative: POV distance and show-don't-tell at
  the emotional peaks.
- At most 5 fixes, ranked by effect on the reader, each with
  where (quoted), what's wrong, why it matters, and the edit
  level it belongs to.
- Label taste as taste. Say what's working.
</requirements>
 
<output_format>
Score table (Criterion | Score | Evidence), the ranked fixes,
then a one-line verdict: another developmental pass, ready
for a line edit, ready for a copyedit, or ready to proofread.
</output_format>
 
<process>
End with a list of the assumptions you made.
</process>

Goals and loops

In an agent that runs commands and checks its own work, such as Claude Code, the Produce stage can run until it measurably passes: /goal keeps starting turns until a condition holds, and /loop repeats a prompt or skill on an interval. The checker reads only the conversation, so every condition below names a check whose output Claude prints, the constraints that must survive, and a bound. How to write your own: Goals and loops.

TaskCompletion condition
Mechanical line edit/goal vale draft.md reports 0 errors and 0 warnings with the project's style (show its summary), the word count stays within 10% of the original, and every change is listed as before → after; or stop after 8 turns
Cut to length/goal wc -w essay.md prints at most 1,200, and every claim in the outline still has a paragraph (print the claim → paragraph map); or stop after 6 turns
Apply the style sheet/goal for every term in style-sheet.md, a search for its banned variants finds 0 matches (print each count), and no sentence was reworded beyond the term; or stop after 6 turns

Loops that fit: rarely: writing doesn't change on its own. The exception is a long document drafted over a session, where /loop 30m can rerun the linter and report only new problems.

What a checker can't judge: voice, argument and whether the piece says what you mean. Keep goals to mechanics; take everything else to the editor's critique in a fresh chat, and never let a goal rewrite your prose.

Prompt library

One entry per deliverable. Quick is the minimum prompt that still works; Full is the skeleton filled in for that deliverable. Replace each [placeholder]: options after a colon are this sheet's vocabulary, so pick one.

Thinking interview

Stage: interview · Use when: you face a blank page, or a topic you know but haven't organized.

Quick

I want to write a [form, e.g. personal essay] about
[topic, e.g. leaving my first job]. Don't write anything:
interview me one question at a time, then give my point in
one sentence. Check: quote me; add no ideas I didn't say.

Full

<context>
Form: [form: essay | op-ed | explainer | report | memo | story]
Reader: [reader, e.g. engineers just promoted to manager]
Topic: [topic, e.g. the first 90 days as a manager]
I have opinions and experiences but no structure yet.
</context>
 
<task>
Interview me so I find what I think before I draft. Don't
write any of the piece.
</task>
 
<requirements>
- One question at a time; follow up on vague answers
  ("what's an example?", "why does that matter to the
  reader?").
- Cover: what I actually think, what I've seen first-hand,
  what the reader gets wrong, the strongest objection, what
  I want the reader to do afterward.
- Quote me; add no ideas, examples or facts I didn't give.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- The thesis is one arguable sentence, in my words where
  possible.
- The examples are ones I told you, quoted.
- The outline follows from my answers, not from a template.
</quality_bar>
 
<output_format>
After about 12 questions: my thesis in one sentence; the
three best examples I gave, quoted; the strongest objection;
an outline of 4-7 points.
</output_format>
 
<process>
Ask the first question now and wait for my answer. After the
summary, end with a list of the assumptions you made about
my reader.
</process>

Voice profile

Stage: produce · Use when: you want edits that respect your style, and a description to paste into every editing prompt.

Quick

Describe my voice from these samples in checkable specifics:
sentence length and variation, how I open and end paragraphs,
favorite constructions, register, humor, what I never do.
Don't praise or improve. Check: every trait has a quote.
[paste 2-5 finished pieces]

Full

<materials>
[paste 2-5 finished pieces of mine, about 2,000 words in all]
</materials>
 
<context>
I'll paste the result into every editing prompt so that line
edits keep my voice. Main forms: [forms, e.g. essays, memos]
</context>
 
<examples>
[paste 1-2 passages of mine I'm proudest of, with why]
[paste 1 passage of mine that misses, with why, e.g. stiff]
</examples>
 
<task>
Write a voice profile: how I write, in specifics an editor
could check a sentence against.
</task>
 
<requirements>
- Sentence length and variation; paragraph length; how I
  open and end paragraphs.
- Diction and register; favorite constructions; punctuation
  habits; kinds of examples and humor; POV and tense habits.
- Things I never do.
- Describe; don't praise, improve or imitate.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
- Under 250 words.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every trait quotes a sentence from my samples.
- Each trait is specific enough to reject an edit
  ("fragments for emphasis, about one per paragraph").
- The profile describes the passages I'm proud of more than
  the one that misses.
</quality_bar>
 
<output_format>
Short sections: Sentences, Paragraphs, Words, Habits, Never.
Then a table: trait | quoted evidence.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made, including traits seen only once.
</process>

Argument stress test

Stage: review · Use when: an argument feels convincing to you and you need to know how it fails.

Quick

Write the strongest case against my argument, as an informed
skeptic would, in about 200 words. Then rank the 3 objections
that do most damage, each with concede, rebut or narrow.
Check: don't soften the objections. [paste argument]

Full

<materials>
[paste the argument or the draft that makes it]
</materials>
 
<context>
Form: [form: op-ed | memo | essay | proposal | grant pitch]
Reader: [reader, e.g. a skeptical board]
</context>
 
<task>
Stress-test the argument before the reader does: map it,
steelman the other side, and rank the objections.
</task>
 
<requirements>
- Map each claim to its evidence and its warrant (Toulmin);
  flag unstated warrants and claims with no evidence.
- A steelman of the opposing case, about 200 words, that a
  smart opponent would sign.
- The 3 objections that do most damage, ranked, each with
  concede, rebut or narrow my claim, and why.
- Add no evidence of your own. Never invent numbers, names,
  quotes or sources: write [NEEDS SOURCE] where support is
  missing.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every objection points at a specific claim, quoted.
- The steelman uses the opponent's best case, not a weak one.
- Recommendations change the argument, not its wording.
</quality_bar>
 
<output_format>
A table: Claim | Evidence | Warrant | Gap. Then the steelman.
Then a table: Objection | Claim it hits | Damage | Concede,
rebut or narrow.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made.
</process>

Line edit

Stage: produce · Use when: structure is settled and you want sentence-level improvement without losing your voice.

Quick

Line edit this passage for clarity, flow and clutter. Keep my
rhythm, fragments and word choices. Return a table: # |
original | suggested | reason; no full rewrite. Check: no
change alters meaning, hedges or numbers. [paste passage]

Full

<materials>
Voice profile: [paste]
Passage: [paste]
</materials>
 
<context>
Form: [form, e.g. a 900-word personal essay]
Reader: [reader, e.g. smart non-specialists]
Structure is settled: don't move or cut paragraphs.
</context>
 
<examples>
[paste 1 line edit you accepted, with why]
[paste 1 edit you rejected, with why, e.g. it cut a fragment]
</examples>
 
<task>
Line edit the passage so no sentence needs a second read and
it still sounds like me.
</task>
 
<requirements>
- Find: sentences a reader would read twice; clutter ("in
  order to", "the fact that"); vague words where a specific
  one exists; logic jumps; runs of same-length sentences.
- Leave alone: [keep, e.g. fragments, the jokes, dialect]
- Each change as small as possible; label taste vs error.
- Never change meaning, facts, quotes or names; never add
  examples or feelings.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
- Don't return the full rewritten text.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every hedge, qualifier and number survives exactly.
- Every suggestion fits the voice profile read aloud.
- No change restructures or copyedits.
- If a sentence works, it's left alone.
</quality_bar>
 
<output_format>
A table: # | original | suggested | reason | taste or error.
Then questions where you suspect I meant something other
than what I wrote.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made.
</process>

Cut to length

Stage: produce · Use when: a piece is over length or feels padded.

Quick

This is [count, e.g. 1,400] words; I need about
[target, e.g. 1,250]. Don't rewrite: list cuts in order of
least loss, each with the exact text and words saved.
Check: every claim and my strongest examples survive.
[paste text]

Full

<materials>
[paste the text]
</materials>
 
<context>
Form: [form, e.g. an op-ed for a regional newspaper]
Now [count, e.g. 1,400] words; target [target, e.g. 1,250]
Limit: [limit: hard cap | soft target | shortest that stays clear]
</context>
 
<task>
List the cuts that bring the piece to length with the least
loss. Don't rewrite anything.
</task>
 
<requirements>
- Cut in this order: redundant sentences, throat-clearing,
  doubled examples, qualifiers that add nothing (not honest
  hedges), clutter phrases.
- Keep every claim, my strongest examples and any setup a
  later payoff needs.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- The cuts add up to the target.
- No cut removes a claim or an honest hedge.
- The through-line survives: every section still serves the
  main point.
</quality_bar>
 
<output_format>
A table: # | delete (exact text) | words saved | why it's
safe, with a running total.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made.
</process>

Copyedit with a style sheet

Stage: produce · Use when: content is final and you need correctness and consistency.

Quick

Copyedit to [style: Chicago | AP | our house style], American
spelling: spelling, grammar, punctuation, usage, consistency.
Mark changes ~~old~~ **new**; query where it depends on my
intent. Check: no quote or style choice is changed.
[paste text]

Full

<materials>
Text: [paste]
House rules: [paste, if any]
Spelled that way on purpose: [paste names and terms]
</materials>
 
<context>
Form: [form, e.g. a 6,000-word report for a nonprofit board]
Content is final; only correctness and consistency are left.
</context>
 
<examples>
[paste a style sheet page you liked, with why]
</examples>
 
<task>
Copyedit the text to [style: Chicago | AP | our house style],
American spelling, and build its style sheet.
</task>
 
<requirements>
- Fix spelling, grammar, punctuation, usage, and consistency
  of names, terms, numbers, capitalization and hyphenation.
- Don't change style, voice, word choice or sentence
  structure unless it is an error; never change quoted
  material or the intentional spellings.
- Where a fix depends on my intent, query as [AQ: ...]
  instead of changing.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every change is an error under the named style, not taste.
- Names, numbers and hyphenation are consistent from first
  page to last.
- Every query is a real ambiguity.
</quality_bar>
 
<output_format>
The text with changes inline as ~~old~~ **new**; the queries
as a numbered list; the style sheet: names and terms as
spelled, number style, hyphenations, rules applied.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made.
</process>

Proofread

Stage: produce · Use when: you're doing the very last pass on final, laid-out text.

Quick

Proofread this final text. List only clear errors: typos,
missing or doubled words, wrong homophones, punctuation,
inconsistent formatting. For each: location, error, fix.
Check: no style suggestions. [paste text]

Full

<materials>
Text: [paste the final text]
Spelled that way on purpose: [paste names and terms]
</materials>
 
<context>
Format: [format: print PDF | web page | slide deck | email]
Everything else is approved; this is the last pass.
</context>
 
<task>
Proofread the text and list every clear error.
</task>
 
<requirements>
- Errors only: typos, missing or doubled words, wrong
  homophones, punctuation, inconsistent formatting, broken
  numbering or cross-references.
- No style suggestions; rewrite nothing; leave the
  intentional spellings alone.
- If a section has no errors, say so.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every entry is an error, not a preference.
- Every location quotes enough words to find it.
- Every section is accounted for.
</quality_bar>
 
<output_format>
A table: # | location (a few surrounding words) | error |
fix.
</output_format>
 
<process>
End with a list of the assumptions you made, such as
spellings you weren't sure were deliberate.
</process>

Fiction scene critique

Stage: review · Use when: a scene or chapter feels flat, slow or confusing and you want to know why.

Quick

Critique this scene without rewriting: goal, conflict and
outcome; POV slips and psychic-distance jumps; telling at
emotional peaks; subtext said aloud. Then the one fix that
matters most, described. Check: quote every line you cite.
[paste scene]

Full

<materials>
[paste the scene or chapter]
</materials>
 
<context>
Genre: [genre, e.g. literary thriller]
POV: [POV: first | close third | distant third | omniscient]
Tense: [tense: past | present]
Want in this scene: [want, e.g. leave before her brother comes]
Placement: [placement, e.g. end of act 1, after the funeral]
</context>
 
<examples>
[paste a short published scene that works, with why]
</examples>
 
<task>
Critique the scene so I know what to fix and why. Don't
write replacement lines.
</task>
 
<requirements>
- Scene (Swain): goal, conflict, disaster; does anything
  change by the end?
- POV: head-hopping, knowledge the POV character couldn't
  have, tense slips outside dialogue.
- Psychic distance (Gardner): where it shifts, and whether
  it suits the moment.
- Show-don't-tell: telling at emotional peaks only, quoted;
  summary and transitions are fine.
- Dialogue: subtext said aloud, "as you know, Bob", voices
  that sound alike; filter words.
- Tension: the line where attention would drift.
- Never invent numbers, names, quotes or sources: write
  [NEEDS SOURCE] where support is missing.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every finding quotes the line.
- Notes diagnose; none supplies new wording.
- The top fix is about the scene's turn, unless the turn
  already works.
- Taste is labeled as taste.
</quality_bar>
 
<output_format>
A table: # | quote | issue | why it matters to the reader.
Then the single most important fix, described in 2-3
sentences.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made about the story outside this scene.
</process>

Worked example

Weak prompt

Make this better.
[pastes a 900-word personal essay]

What's wrong

  • No level of edit, so the model blends all four and rewrites every sentence.
  • No reader or purpose, so "better" means the model's average idea of good prose.
  • No voice description or limits: the deliberate fragments get smoothed and honest hedges removed.
  • No ban on invention: it adds a vivid anecdote that never happened and inflates the ending into a lesson.
  • No change record: clean text comes back, and the author can't tell which ideas are still theirs.

Strong prompt

<materials>
[paste the essay]
</materials>
 
<context>
A 900-word personal essay for a general-interest magazine;
readers are smart non-specialists. Structure is settled:
don't move or cut paragraphs. My voice: short declarative
sentences broken by long ones; dry understatement; fragments
for emphasis; no exclamation marks.
</context>
 
<task>
Line edit the essay so no sentence needs a second read,
keeping my voice.
</task>
 
<requirements>
- Find: sentences a reader would read twice, clutter, vague
  words where a specific one exists, logic jumps.
- Keep the fragments, the understatement and every hedge.
- Add no facts, examples or feelings. Never invent numbers,
  names, quotes or sources: write [NEEDS SOURCE] where
  support is missing.
- Each change as small as possible; label taste vs error.
- Don't return the full text.
</requirements>
 
<quality_bar>
Before answering, check your draft against:
- Every hedge, qualifier and number survives exactly.
- Each suggestion sounds like the voice described above.
- No change restructures or adds content.
</quality_bar>
 
<output_format>
A table: # | original | suggested | reason | taste or error.
Then three things that are working.
</output_format>
 
<process>
Ask me up to 3 questions first if anything missing would
change your answer; otherwise go on. End with a list of the
assumptions you made.
</process>

What came back

The response, illustrative, abridged:

1. "In order to understand why I left" -> "To understand why
   I left". Clutter. Error.
2. "She was, in a way, sort of the reason." -> "She was, in
   a way, the reason." Doubled hedge; keeps "in a way".
   Taste.
3. "Gone. All of it." No change: the fragment lands the
   paragraph.
4. "I may have been wrong about him." -> "I was wrong about
   him." Hedging weakens the turn. Taste.
5. "the day of the thing with the car" -> "the day of the
   accident". Vague noun. Taste.
...
Working: the opening image of the kitchen; the restraint in
paragraph 6; the last line, which doesn't explain itself.
Assumptions: "him" in paragraph 4 is the father; the essay
is in past tense throughout on purpose.

How an expert judged it

Against the <quality_bar>:

  • Level respected: every change is sentence-level; nothing moved or added. Pass.
  • Voice: row 3 leaves the fragment alone and row 2 keeps one hedge on purpose. Pass.
  • The flaw: row 4 deletes an honest hedge. "I may have been wrong" becomes "I was wrong": a doubt becomes a verdict. That fails the first check (every hedge survives) and the "meaning preserved" criterion, and labeling it taste hides that it changes what the author claims. In a personal essay the narrator's uncertainty may be the point. Reject it (stet).
  • Row 5 needs the author, not a rule: "the thing with the car" may be the dry understatement the brief asked to keep. Labeled taste correctly; the author decides.
  • Verdict: accept 1 and 2, stet 4, decide 5, then check the rest for the same kind of shift before going on.

The follow-up

Stet #4: "may have" is deliberate; the essay turns on not
knowing. Before you continue, list any other suggestion that
changes certainty, emphasis or a hedge, quoting original and
suggestion. Then continue the line edit from paragraph 5
with the same rules, and don't re-suggest #4.

Judge the output

CriterionGoodRed flag
Level respectedonly the requested kind of changea line edit that restructures; a proofread that rewrites
Diagnosis firsteach note quotes the passage and says why it matters to the reader"consider revising for clarity" with no location
Voice intactread aloud, it still sounds like youyour fragments, jokes and odd word choices smoothed away
Meaning preservedevery claim, hedge and number survives"may" became "will"; a qualifier vanished
Nothing inventedno new facts, anecdotes, quotes or sourcesa vivid detail you never wrote
Prioritizedthe biggest problems first40 equal-weight nitpicks
Honestnames real weaknessespraise for everything (models tend to flatter)
Taste labeledseparates preference from errorpresents style opinions as rules

The same checks, condensed, are in every <quality_bar> on this sheet; the full scored critique is stage 4 of the chain. Follow-ups for the moments in between:

Below are an ORIGINAL and an EDITED version. List every
claim, number, qualifier and hedge in the original, and say
whether the edit keeps it exactly, weakens it, strengthens it
or drops it, quoting both. Then list anything in the edit
that isn't in the original.
ORIGINAL: [paste]
EDITED: [paste]
Be harsher. Assume this draft will be rejected. What is the
most likely reason an editor at
[publication, e.g. the magazine you're pitching] would give?
Quote the passages that would cause it.
Read this as [reader, e.g. a smart 16-year-old] who knows
nothing about [topic, e.g. index funds]. Paragraph by
paragraph: what you understood, the first word or step you
didn't, and the question you'd ask. Don't explain or fix.
[paste text]
Which of your suggestions are matters of taste rather than
clarity or correctness? Remove those and show me what's left.

Pitfalls

PitfallWhy it happensFix
"Make it better" rewrites everythingno level, no limitsname the level; ask for a change table
Voice flatteningthe model pulls toward average diction and rhythmvoice description, editor rules, smallest possible changes
Flatterymodels tend to agree and praiseask for the three biggest problems; ask it to argue the piece fails
Invented facts, quotes and sourcesfluent completion of what "should" be there"use only what I gave you"; fact-check every addition
Wrong orderline editing a draft that needs restructuringdevelopmental first; copyedit and proofread last
Fiction conflicts resolved, subtext explainedthe model's helpful defaultscritique only; you write the lines
Accepting the first ideathe first suggestion is the most predictableask for many options; discard the obvious ones
Losing track of changesclean rewritten text returnedtracked-change output, or Compare documents
Long manuscripts in one pastelong inputs get uneven attention, and context windows have limitschapter by chapter, with a running synopsis and style sheet
Copyedit errors in names and termsthe model "corrects" unusual spellingsgive it a list of intentional spellings; review every change
Using a model as a sensitivity readerit can't supply lived experiencehire readers from the community portrayed
Your style drifting toward the model'smonths of accepting its phrasingdraft first yourself; reread your pre-AI writing now and then

References