Margin & Method
Prompt Craft

AI Editing Prompts for Four Separate Passes

Margin & Method · Updated 2026-07-19 · 7 min read
AI Editing Prompts for Four Separate Passes
In shortUseful AI editing prompts assign one pass at a time: developmental, line, copy, or proof. Give the audience, purpose, house rules, protected facts or quotations, and a strict output format. Ask the tool to quote the passage behind each suggestion and explain the reason before rewriting. Review every change manually, because a detailed prompt can constrain the task but cannot guarantee factual accuracy, style compliance, or good judgment.
Contents

Prompt for one editorial pass at a time

The most useful AI editing prompts name one pass—developmental, line, copy, or proof—and define the audience, purpose, protected material, and output format. Ask the tool to quote the passage behind each suggestion and explain the reason before rewriting. Then review every proposed change manually. A detailed prompt can narrow the task; it cannot guarantee accuracy, taste, or obedience when the model encounters an inviting sentence and develops renovation plans.

Use the four templates below as protocols, not incantations. Tool behavior varies across models, versions, settings, accounts, and dates. Test each prompt on representative copy, inspect what it misses, and revise the instructions for your workflow.

Prepare a reusable context block

Put the stable facts at the top of every pass:

DOCUMENT PURPOSE:
AUDIENCE:
DESIRED READER ACTION OR UNDERSTANDING:
PUBLICATION VOICE:
HOUSE STYLE RULES RELEVANT TO THIS PASS:
PROTECTED FACTS, FIGURES, NAMES, QUOTATIONS, LINKS, AND TERMS:
MATERIAL THE TOOL MUST NOT ADD OR INFER:
OUTPUT FORMAT:
TEXT BEGINS:
[paste text]
TEXT ENDS

Fill the fields. An empty label is not a constraint; it is office furniture.

Place the text between unmistakable boundaries so instructions quoted inside the draft are treated as content. Do not include confidential or personal information unless the system and use are approved. Check the provider's current documentation and the rules governing the work before uploading any draft.

Protected material deserves a practical handling rule. If wording must remain exact—an approved quotation, legal clause, product name, or safety instruction—lock it outside the generated rewrite when possible. Telling a model “never change this” lowers risk but does not create certainty.

Pass one: developmental diagnosis

Developmental editing asks whether the piece has the right purpose, argument, order, scope, evidence, and proportion. Do not request polished replacement prose yet.

Act as a developmental editing assistant. Diagnose this draft for the
stated audience and purpose. Do not rewrite it.

Return a table with these columns:
1. section or location
2. exact quotation from the draft
3. issue type
4. reader impact
5. question the editor should answer
6. possible structural action, without replacement prose

Check only for: buried main answer, missing context, unsupported logical
leap, weak order, repetition, scope drift, uneven section weight, and an
ending that fails to complete the stated purpose.

Do not invent missing facts, sources, examples, quotations, or audience
research. If evidence is needed, label it NEEDS EVIDENCE. If the quoted
passage does not support a concern, omit the concern.

Why the quotation column matters: it forces the criticism to attach to real text. Why the editor question matters: some apparent defects are unresolved choices, not sentences waiting for machine repair.

Review the table before changing the draft. Reject observations that conflict with the brief, misread deliberate repetition, or import a generic article structure. Select perhaps three to five structural issues for the next revision. A report with 34 concerns may be comprehensive in the way tipping a filing cabinet onto the floor is comprehensive.

Pass two: line editing with voice protection

Line editing improves clarity, flow, emphasis, rhythm, and economy within an approved structure. It is where generic “sound professional” instructions do the most damage.

Act as a line editing assistant. Preserve the author's point of view,
register, and deliberate humor. Do not change facts, figures, names,
quotations, citations, links, or the strength of a claim.

Flag no more than 12 high-value line edits. Return a table with:
1. exact original sentence or passage
2. issue: clarity, rhythm, emphasis, redundancy, ambiguity, or tone
3. concise explanation tied to the audience
4. proposed revision
5. meaning-change risk: low, medium, or high

Do not standardize dialect, unusual but clear syntax, fragments used for
effect, or specialist terms the audience knows. Do not add metaphors,
examples, transitions, or claims. If a passage may have two meanings,
ask a question instead of choosing one.

Cap the number of changes. Without a limit, the tool may keep editing because output is its only hobby. A cap encourages prioritization and makes human review possible.

Check every proposal against the original meaning. Watch modal verbs and qualifiers: may, can, often, some, and under these conditions carry real information. A smoother universal statement can be materially false.

For more control, split the pass again. Ask first for ambiguity and redundancy, approve those items, and only then request rhythm or tone suggestions. The human-owned editing workflow explains how to version and approve those decisions.

Pass three: copyediting against named rules

Copyediting needs an actual standard. Supply the publication dictionary, spelling convention, capitalization list, number style, punctuation preferences, terminology, and any accessibility rules relevant to the piece.

Act as a copyediting assistant. Apply only the supplied house rules and
standard grammar needed for clear meaning. Do not perform developmental
or stylistic rewriting.

Return a change table with:
1. location
2. exact original
3. proposed change
4. rule or reason
5. confidence: high, medium, or query

Check grammar, agreement, punctuation, spelling, capitalization,
hyphenation, number treatment, heading consistency, defined acronyms,
parallel lists, and inconsistent terminology.

Do not alter quotations. Do not claim a house rule that was not supplied.
Put uncertain cases in a separate QUERY list. Do not silently resolve
factual conflicts or ambiguous meaning.

The rule-or-reason column exposes invented authority. If the explanation says “for consistency” but the term appears only once, the suggestion has not earned its chair.

Run a separate factual review for names, dates, statistics, citations, product behavior, and external claims. Copyediting can notice that two dates conflict; it cannot determine which date is true without reliable evidence. Use the claim-ledger workflow for that job.

Pass four: proof the stable version

Proofreading happens after content, facts, and layout are nearly final. Limit the tool to surface defects and ask it not to improve style.

Proofread the supplied final copy. Do not rewrite for style, tone,
conciseness, or persuasion. Do not add facts or alter quotations.

Report only:
- typographical errors
- duplicated or missing words
- punctuation or spacing errors
- inconsistent heading levels
- broken numbering or list sequence
- caption or label mismatches visible in the supplied text
- visibly malformed link text

Return: location | exact error | correction | reason.
If no issue is found in a category, say NONE. Do not manufacture a change
to demonstrate activity.

Proof the rendered artifact as well as the source text. A text model cannot reliably assess line breaks, image crops, visual hierarchy, hidden overflow, or a button that has wandered under the footer unless the relevant rendered state is actually available to it. Human visual review remains part of the release.

Add a no-new-claims gate

After any pass that produces replacement wording, compare the proposal with the source and ask:

If yes, reject the change or open a fact-check row. Asking the model to run this gate can help surface candidates, but it cannot be the final judge of its own invention. A smoke alarm is useful; it does not sign the fire inspection.

Ask for questions instead of guesses

Add this clause when the source is incomplete:

When a responsible edit depends on information not present in the draft,
ask a numbered question. Do not supply the missing information, create a
generic example, or infer the publication's decision.

This is especially important for voice, audience, legal meaning, medical or safety context, source intent, and quotations. A visible question preserves the uncertainty for the person who can resolve it.

Evaluate prompts with a small test set

Build a test set of five to ten short passages that represent the work: a clean paragraph, a buried answer, a deliberate fragment, an unsupported statistic, a quotation, an ambiguous pronoun, and a house-style exception. Record what a good response should flag and what it must leave alone.

Run the prompt, then score:

Repeat after a model or workflow change. Do not rely on a prompt merely because it behaved last month; neither tools nor the surrounding product settings are fixed furniture.

Stop when another pass creates churn

More passes do not automatically produce more quality. Stop when suggestions repeat, swap one acceptable phrase for another, flatten distinctive voice, or cost more review time than the remaining issues justify.

If the assisted work is material, apply the publication rules and the AI-use disclosure guide before release.

The sequence is simple: diagnose structure, revise approved lines, apply named copy rules, and proof the stable version. Between every stage, a human chooses. Prompts make those choices easier to inspect. They do not make them optional.

FAQ

What should an AI editing prompt include?

Include the document's purpose, audience, editing level, house rules, non-negotiable facts, protected quotations, desired output format, and explicit exclusions. Provide the text in clear boundaries. Ask for quoted evidence and reasons. If the task cannot be described precisely, the tool cannot rescue it by guessing more confidently.

Why separate developmental editing from copyediting?

They answer different questions. Developmental editing examines purpose, argument, order, gaps, and proportion; copyediting examines correctness and consistency at sentence level. Combining them encourages scattered output and premature polishing. Resolve the large decisions first, freeze the new structure, and only then spend attention on commas that may survive the renovation.

Can a prompt prevent an AI tool from changing facts?

A prompt can tell the tool not to alter facts, figures, names, quotations, or citations, but the instruction is not a guarantee. Models can still paraphrase inaccurately or introduce material. Compare every suggestion with the source draft and verify claims independently. Protected text should be locked or handled outside the generated rewrite when possible.

Should I ask for a rewritten article or an edit table?

Use an edit table first when traceability matters. Request the location, quoted original, issue type, reason, and proposed change. That lets the editor reject weak suggestions without searching a transformed draft. A clean rewrite can follow after decisions are approved, but it should be compared against the original and the accepted-change list.

How many editing prompts should I run?

Run only the passes the document needs. Four passes are a useful menu, not a compulsory obstacle course. A short routine update may need copy and proof checks; a new argument may need developmental and fact work first. Stop when additional passes produce stylistic churn rather than clearer, more accurate writing.

Do the same prompts work in every AI writing tool?

No. Models, versions, context limits, system instructions, file handling, and product features vary and change. Treat each prompt as a testable protocol, not a magic incantation. Try it on representative text, inspect failures, revise the instructions, and confirm current provider documentation before relying on any tool-specific behavior.