This page is the full tutorial, with every prompt and some extra exercises to work through at your own pace. You can also view the presentation slides.

Teach AI how to read your work. You stay the writer and the editor.

Introduction

Why editing is the place to start

Many people first use these tools for something they can’t check, like a summary of a literature they haven’t read or a draft in an unfamiliar genre. It’s hard to tell a good answer from a plausible one there.

Editing is different, because you know what you meant your paragraph to say. You can hold any comment or proposed edit up against your own text and decide whether it helps. That makes editing a good place to learn what these tools do well: ask for a reading first, and revise once you’ve seen its reasons.

What you’re working with

You don’t need to know how a model works inside to judge what it tells you. Four habits of these tools come up in the exercises:

  • It edits smoothly, and settles ambiguities without saying so. A cleaner sentence can quietly answer a question you meant to leave open. (Step 1)
  • It answers the question you ask. An open question like what am I missing? brings in broad advice, which is often worth having. Asking it to quote your text gets findings you can check sentence by sentence. (Step 2)
  • It will read as anyone you name. A student and a reviewer notice different things, but the same model wrote both readings, so they don’t count as two opinions. (Step 3)
  • It forgets your preferences. Each new conversation starts from nothing unless you save your rules and bring them along. (Steps 4 to 6)

Before you start

  • A laptop, with an AI assistant open in a browser tab. Claude, ChatGPT, Gemini, or Copilot all work, and a free account is fine.
  • One paragraph of your own writing, such as a draft, an abstract, a policy, or a grant blurb. Your own work teaches more than an invented example. No paragraph? Take one from the practice texts.

Use texts you have permission to share, such as your own writing or the practice texts.

The paragraph we’ll use

We begin with the AI-use policy from the practice texts. It is short enough to read together and familiar enough to judge. For the style-guide exercise, we move to a four-paragraph project report, where wording choices have to hold across a longer argument.

Use of AI Tools. Students may use generative AI tools in this course where appropriate, but should not use them to complete assignments for them. Any use of AI should be disclosed. Work that is substantially produced by AI will not receive credit and may be referred as an academic integrity violation. Students are responsible for the accuracy of everything they submit. If you are unsure whether a particular use is permitted, ask before submitting. The goal of this policy is to support learning rather than to prohibit technology, and I expect students to use their judgment.

Read it as the person who wrote it and it seems clear. This session is about reading it as the students it is addressed to.

About the replies on this page. Recorded replies show the tool and date. They come from the conversation in which this workshop was developed, so the tool already had some context; re-run them in a fresh conversation before treating them as typical. […] marks a cut. Examples labeled prepared were written for the slides rather than recorded. Your own replies will differ, and a different reply can be just as good.

1. Try a reasonable editing request

Paste the syllabus policy, then ask for a conventional edit.

A reasonable first request
Please edit this syllabus policy for clarity and concision. Preserve my meaning and tone, and explain any substantive changes.

Paste the syllabus policy, then send this. Save the reply and compare its changes with the original. For another text, replace 'syllabus policy' with its genre.

Some of what comes back will probably help. Read the reply beside the original and pick one change you would accept and one you would want to argue with.

  • Did “should,” “may,” and “must” keep their different meanings?
  • Did the edit make a sentence clearer, or did it settle a policy question the writer had deliberately left open?
  • Did it explain the changes that alter meaning, or only the small ones?
  • Which questions are still unanswered now that every sentence reads smoothly?

“Preserve my meaning” only helps if a sentence has one meaning to begin with. Where the original is ambiguous, the tool either picks a reading or asks which you meant, and a question is the better result. Step 2 gets it to name those decisions before it rewrites anything.

Add one line to any editing request: “If a sentence could mean more than one thing, ask me which I meant before you change it.”

Prepared example: “Students may use AI where appropriate, but not to complete assignments for them.” The sentence is shorter, but “appropriate” still needs examples, and the paragraph still doesn’t say how to disclose AI use.

Now send the same request with your own paragraph, and look for a choice the edit made for you.

2. Teach it to read like you do

Use the same original policy, this time with a specific job. Name the reader, focus the concern, and ask for the evidence behind each observation.

Give the policy reader a specific job
This syllabus policy is for first-year students.

Before editing, identify unclear instructions. Focus on disclosure.

Quote the passage, explain what a student cannot do, and name the decision I need to make. Don't invent a policy.

Use the same original policy as in the general editing request. Compare the usefulness of the observations, rather than assuming that a longer prompt always works better.

Prepared example: “Any use of AI should be disclosed” doesn’t say where or how. The instructor needs to decide what disclosure includes and where it goes, then add that to the policy.

That reply gives you a passage, a problem, and a decision to make, because the prompt asked for a diagnosis before any editing.

For other documents, use the general brief below. A brief says what the text is, who reads it, what it’s trying to do, what worries you, and what shouldn’t change, and it asks for comments rather than a rewrite. The recorded reply after it comes from running this brief on the policy.

The editorial brief
This is a [genre: seminar paper / abstract / syllabus policy / cover letter] for [audience]. It is trying to [do what].

I am worried about [first concern] and [second concern]. Do not change my [voice / hedging / terminology / citation style].

Don't rewrite anything. Read it and tell me the three places a reader is most likely to get lost. Quote the exact sentence, and say why it loses them.

Fill in the brackets. The last line is the important one: you are asking for a reading, not a rewrite.

The same paragraph, briefed Claude Opus 5 · 2026-09-22

Three places a reader is most likely to get lost:

1. > “Students may use generative AI tools in this course where appropriate”

“Where appropriate” is the whole policy in three words, and it is the three words that don’t say anything. A student who wants to comply has no way to determine appropriateness from this sentence, and the paragraph never comes back to define it. Everything after this point assumes the reader already knows.

2. > “Work that is substantially produced by AI will not receive credit”

“Substantially” is a threshold with no number, no proportion and no example. A student who used AI to reorganise an outline and then wrote every sentence themselves cannot tell which side of this they are on — and on this text, neither can you. It is also the sentence with the penalty attached, so it is where the ambiguity costs most.

3. > “Any use of AI should be disclosed.”

Disclosed how, to whom, in what form? A footnote, a cover note, an appendix listing prompts? The sentence creates an obligation and specifies no way to discharge it, so a conscientious student either over-discloses or guesses.

On your two worries: enforceability and disclosure are the same problem here. Both rest on “substantially,” and until that word names a test, neither you nor the student can apply it consistently.

Brief filled in as: a syllabus policy, for undergraduates in a history course, trying to set expectations students can actually follow; worried about enforceability and about what counts as disclosure; don't change voice or terminology.

Compare this with your first edit and look for a question the edit didn’t raise. To compare prompts fairly, use the same text and tool in fresh conversations and keep both replies.

Open questions and anchored questions

Both kinds are worth asking. An open question, such as anything else I should worry about? or what am I missing?, lets the tool bring in what you didn’t think to ask about: teaching practice, the conventions of a genre, how readers like yours tend to react. An anchored question ties each answer to a passage, so you can check it against your text. Ask open questions to widen what you consider, and anchored ones when you need to know what is on the page.

The two are easy to confuse in a reply. Later in the same conversation, an open follow-up produced this:

Try an open-ended follow-up
Anything else I should worry about?

After a reading of the policy, send this broad follow-up. Open questions bring in things you hadn't thought to ask about. Then sort the reply: which points are about your text, and which are general advice?

A comment that sounds like a finding Claude Opus 5 · 2026-09-22

The tone is generally supportive, but the policy could do more to convey the pedagogical rationale behind these restrictions. Students respond better to guidelines when they understand the reasoning, and there’s an opportunity here to frame AI use as a skill to be developed rather than a risk to be managed. The final sentence gestures at this but could be developed further.

Follow-up question, asked openly: “anything else I should worry about?” The briefed prompt had asked for quotations and this one didn't, so the reply mixes a point about the policy with general advice.

Some of this is good advice. Explaining the reasoning behind a policy and treating AI use as a skill are both worth considering. But the reply reads like a diagnosis of the paragraph, and you can’t tell which parts are. Ask it to anchor what it said:

Make it show its work
Quote the exact words that made you say that. If you can't point to specific text, say so instead of explaining further.

Use this when you need to know whether a comment is about your text. If it can't point to a passage, treat the comment as general advice: it may still be worth taking, but it isn't a finding about this draft.

The same comment, made to quote Claude Opus 5 · 2026-09-22

I can point to one of the three:

“The goal of this policy is to support learning rather than to prohibit technology”

That is the sentence I meant by “gestures at this.”

For the other two I can’t. “Students respond better to guidelines when they understand the reasoning” is a general belief about teaching, not something I read in your text. And “an opportunity to frame AI use as a skill to be developed” describes a policy you did not write rather than a feature of the one you did. Neither was a finding about your paragraph.

It tied one point to a sentence in the policy and identified the other two as general advice. One sentence was enough to sort the reply.

Now you can sort the reply. One point is about a sentence in the policy, and you can go and check it. The other two are general advice: decide whether you agree, but don’t treat them as problems with this policy.

Ask open questions when you want ideas, and ask for the exact words when you want to know what’s on the page. A comment that can’t point to a passage is general advice, which may still be worth taking.

A quotation doesn’t prove the criticism is right, but it gives you somewhere to look. When the claim is that something is missing, ask which purpose or rule makes the absence matter.

These tools also tend to open with praise. When you want criticism, ask in a form where something has to come last:

Turn off the flattery
Don't tell me what's working. Rank my paragraphs from weakest to strongest and explain what makes the weakest one weak.

A yes-or-no question like "is this good?" usually gets a yes. A ranking has to put something last, which gives you a place to start.

Which instructions to keep

The brief changes with every document, so write it fresh each time. That’s worth doing anyway, because writing it is how you work out what the piece is for.

Two other instructions describe how you want any of your writing read: quote the text when you comment on it, and label general advice as general, and don’t open with praise. Type them by hand today, then make them the first lines of the rule file you build in step 5. Step 4 adds the rules you can only find by watching the tool get your own work wrong.

3. Assign reader roles

Keep the text and the brief fixed, and change only the reader. A student asks what to do next. An assignment designer asks whether the instructions and the grading criteria describe the same piece of work. Neither has to rewrite anything to give the author a question worth answering.

Read a real assignment in two roles

Open the Making History assignment exercise. It includes the actual Campus History archive-and-essay assignment, a text copy, and links to its surrounding instructions. Give the same materials to both readers in separate conversations.

Read the assignment as a student
Read this assignment as a first-year student encountering it for the first time. Do not complete the assignment or rewrite the instructions.

Explain what I would do first, what I think I must submit, and how I would judge whether my work meets the expectations. Quote the instructions that support each part of your reading.

Identify up to three places where I would have to guess. Separate what the assignment says from what you inferred. Treat information absent from these materials as a question for the instructor, not proof that it was never explained elsewhere.

Use the complete Campus History Part 2 assignment. Supply any course schedule or rubric you want included, and list what you supplied.

Read the assignment as a designer
Review this assignment using transparent assignment design:
- Purpose: what students will learn and why it matters.
- Task: what they must do, in what sequence, and submit.
- Criteria: how they can recognize successful work.

For each, quote the evidence and distinguish what is clear, partly explained, or not stated in the supplied materials. Recognize strengths as well as gaps. Check whether every graded component is introduced in the task instructions.

Prioritize three questions or small clarifications. Do not invent requirements, dates, grading weights, or teaching activities. Do not treat missing section labels as proof of poor design. Do not rewrite the assignment or give it a numerical score.

Use the same assignment and supporting materials in a separate conversation. The review uses TILT's purpose, task, and criteria framework.

The designer’s framework is purpose, task, and criteria for success, from TILT. Give what the assignment already explains as much attention as what may be missing.

Compare the student and designer readings
Compare these two readings of the same assignment.

Where does a student's practical uncertainty connect to a design issue? Where do the readings differ? Check each claim against the assignment and quote the evidence. Identify any claim the readings do not support.

Recommend one clarification for the instructor to consider, then state what decision the instructor still needs to make. Do not resolve that decision or rewrite the assignment yet.

Supply the assignment and both full readings. Agreement between generated roles is a lead to investigate, not independent confirmation.

Prepared observation, in TILT’s three parts:

  • Purpose. “Skills you are practicing” explains what students learn and why.
  • Task. The archive visit and the essay requirements give a workable sequence.
  • Criteria. “What I’m looking for” describes strong work. But the grading also expects an “AI-Archive Comparison” that the requirements never place, so students can’t tell where it goes or whether it counts toward the word target.

A student would ask where to put the comparison, and a designer would ask why the grading expects something the instructions never assign. Both lead to the same missing sentence. Recorded runs of all three prompts are on the exercise page. The designer’s reading marks the comparison “No. It appears only in the criteria,” and the comparison then calls that overstated: the task does ask students to compare, but never names the comparison as something to hand in.

The instructor decides how to fix it, and other course materials or class discussion may already cover it. Before you treat an AI reading as evidence of what students understand, ask some students to explain the assignment back to you.

Try other readers on your own text

You already imagine readers when you revise: the reviewer who may object, the student meeting a term for the first time, the editor in a hurry. Name that reader in the prompt and see what it helps you notice.

Run at least two roles on your paragraph, and ask for comments only, not fixes.

Reader — the skeptical peer reviewer
Read this as a skeptical peer reviewer in [field] who is inclined to recommend major revisions. Don't fix anything. Tell me the single claim you'd push back on hardest, and the question you'd put in your report.
Reader — the smart outsider
Read this as a smart graduate student from a different discipline. Don't fix anything. Tell me every point where you had to guess what I meant, and what you guessed.
Reader — the tired panelist
Read this as a grant panelist who has 40 applications to get through and four minutes for mine. Don't fix anything. Tell me what you took away, and where you started skimming.

Then add the one closest to the work you brought:

Reader — the desk editor (for researchers)
Here is a journal's aims and scope, then my abstract. Read as the desk editor. Would you send this out for review or desk-reject it? Give me the grounds, in the journal's own terms.

Paste your abstract plus the journal's aims-and-scope statement.

Reader — the confused student (for teaching)
Here is an assignment I give to undergraduates. Read it as a sophomore who wants a good grade and is reading it for the first time, the night before it's due.

Don't improve it. Tell me: what do you think I'm asking for, what would you do first, and what are you unsure about?

Paste an assignment prompt instead of your own prose. This one tends to be uncomfortable in a useful way.

Reader — the grader (for students)
Here is the rubric for an assignment, then my draft. Read it only against these criteria. For each criterion, tell me where my draft currently meets it and where it doesn't, quoting the relevant text. Don't rewrite anything and don't give me a grade.

Paste the rubric, then your draft. It asks for a diagnosis, and the words stay yours.

Here is how the confused-student role read the syllabus policy:

Read as a confused student at 11pm Claude Opus 5 · 2026-09-22

What I think you’re asking for: I can use AI but not too much? I think I’m allowed to use it for some things as long as I say I did. But I’m not sure if that means I should just avoid it to be safe.

What I’d do first: Honestly, probably nothing — I’d not use it at all, because the academic integrity part is scary and it’s late and I don’t want to risk it. If I’d already used it to help me outline, I’d panic a bit and try to work out whether that counts.

What I’m unsure about:

  1. What counts as “appropriate”? Is using it to fix my grammar appropriate? To explain a reading I didn’t understand? Those feel different to me but the policy treats them the same.
  2. What does “substantially produced” mean? If I wrote every sentence but asked it to reorganise my paragraphs, is that substantial? Half my words? Any of my words?
  3. How do I disclose? Do I write it at the end, email you, is there a form? And if I disclose something that turns out not to be allowed, have I just reported myself?
  4. If I ask you beforehand, does that go on a record anywhere?

That last one is the one I’d actually get stuck on. The policy says to ask if I’m unsure, but asking means telling my professor I was thinking about using AI, and I don’t know how that lands.

The prompt says “assignment”; it was adapted to “syllabus policy” and otherwise sent as written. It asked four questions, as sessions/editorial-assistant/practice-texts.md says it would, and the fourth is the one to read closely.

And here is the outsider’s reading of the same policy:

Read as a smart outsider Claude Opus 5 · 2026-09-22

Points where I had to guess:

  1. “where appropriate” — I guessed it means “for tasks that support your learning rather than replace it,” because the last sentence implies that. The policy never says it, so I may have imported it.
  2. “complete assignments for them” — I guessed this means producing submitted text. It could equally mean doing the thinking: having AI settle your argument and then writing it up in your own words. Those are very different, and a literal parse covers only the first.
  3. “substantially produced” — I guessed a proportion of words, maybe more than half. But it could mean the intellectual contribution regardless of word count. The two tests would sort different students differently.
  4. “should be disclosed” — I guessed mandatory, from context, though “should” is the weakest modal in the paragraph while everything around it uses “will” or “may.”
  5. “may be referred” — I guessed “at the instructor’s discretion” rather than “happens automatically.” That is a large difference if you are the student.

The one I am least sure of is 2, because it decides whether the policy is about labour or about authorship, and those produce different rules.

The same paragraph read calmly instead of anxiously. It stops at the same words and calls the problem something else: a definition problem rather than a risk.

Compare the readings

Put the readings side by side. Then check the comparison against both replies and the original, since a tidy summary can still misrepresent all three.

Compare the readers
Compare those readings. Where did the two readers disagree about what the problem was? Don't resolve the disagreement — just tell me what it suggests I haven't decided yet.

Where the readers disagree, look for a decision you haven't made yet.

The two readings, side by side Claude Opus 5 · 2026-09-22

Where they disagreed:

On what kind of problem this is. The student reads the policy as a risk to be navigated — every question is about exposure, ending with “have I just reported myself?” The outsider reads it as a definition problem: which test does “substantially” name. They stop at the same sentences and call the trouble something different. The student never flags “may be referred” as ambiguous; the outsider never registers that asking is itself costly.

On “ask before submitting.” The policy offers this as the remedy. The outsider accepts it as one. The student identifies it as a fifth problem, because using it means disclosing an intention.

What that suggests you have not decided: whether this policy is instruction or enforcement. If it is instruction, the student’s reading is the one that matters and the fix is worked examples of appropriate use. If it is enforcement, the outsider’s reading matters and the fix is defining “substantially” as a test you could apply to two students and get the same answer. The present text is trying to be both, which is why “ask before submitting” reads as an open door to one reader and a tripwire to the other.

Both readers stopped at “substantially.” They disagreed about what kind of problem it is, which points to a decision the author hasn't made.

Where the readings agree, you have learned where to look. You have not learned that the criticism is right.

Read the comparison for three things.

What both readers flagged. Go back to the passage and see whether it supports the concern. In the assignment, the comparison requirement really is named in the grading section. Where it belongs is still the author’s call.

What only one reader flagged. Decide whether that concern matters for the audience you actually have. A specialist asking for more detail and a newcomer asking for a definition can both be right, for different readers.

Where the readers disagree. Work out what the disagreement is about: a choice you made on purpose, a misreading, or context you left out. In the policy comparison above, the reply offers “instruction or enforcement” as though you had to pick one. A policy can do both. Keep the question it raises and drop the either-or.

Use the questions that hold up to plan a revision, or to decide what to ask a real reader.

Give each imagined reader its own conversation and ask for questions, not fixes. Check the concerns they share against the passage, and the ones that matter against a real reader.

4. Turn corrections into rules

The first three steps gave you changes and comments to judge. Some of your corrections will apply beyond this policy: keep a qualification, keep a defined term, ask before resolving an ambiguity. Write those down as rules you can reuse. Whether a rule gets followed depends mostly on how it’s phrased.

Start from your corrections

Don’t start with a blank page. A style sheet written from scratch tends to describe the writer you’d like to be, which doesn’t work as an instruction. Your corrections are better material: each time you said no, put that back, you found a possible rule.

Find the rule you keep applying
Look back over this conversation. I corrected you several times.

For each correction, tell me what rule I was applying — written as an instruction you could follow next time, not as a description of what I seemed to prefer.

Quote the exchange each rule comes from. If you can't point to a specific correction, don't list the rule.

Run this at the end of a working session, while the corrections are still in the transcript. It describes your own judgment back to you, so expect to disagree with part of it.

If you’ve made the same correction twice, make it a rule, so you don’t have to fix it by hand a third time.

A rule has to be checkable

Test each rule with one question: could someone who doesn’t know you tell whether the rule was followed, just by looking at the text?

“Be more scholarly” fails, because there’s no clear way to break it. “Never replace a hedged verb with an unhedged one” passes: if suggests becomes demonstrates, you can point at the change. Most first drafts of a rule name a quality (rigorous, clear, academic) where they need to name an action you could find in the text.

Two more things to build in:

  • Say what not to do. “Write clearly” is satisfied by anything the tool considers clear, which usually means its own style. Say what you want, then rule out the way it usually goes wrong.
  • Put your rules in order. “Never remove my hedges” and “cut anything that isn’t working” will eventually apply to the same sentence. If you haven’t said which one wins, the tool may decide differently each time. Five rules in a clear order work better than twenty with none.
Make a rule checkable
Here is a rule I wrote for my style sheet:

[paste the rule]

Could someone who doesn't know me tell whether this rule was followed, looking only at the text? If not, rewrite it so they could — as a specific action to take or to avoid.

If it can't be made checkable, say so, and tell me what I would have to decide first.

The test every rule has to pass. Most first drafts name a quality rather than an action, and this is what catches it.

Then test the rule on a passage where it shouldn’t apply, to make sure it doesn’t fire everywhere.

Test the rule where it shouldn't fire
Here is one of my rules, then three passages.

[paste the rule]
[paste three passages — at least one the rule should NOT apply to]

For each passage: does the rule apply? Quote the exact text that triggers it, or say that the rule doesn't apply here.

I am looking for cases where the rule fires when it shouldn't. Tell me if you find one.

A rule that catches everything is as useless as one that catches nothing. Include at least one passage the rule should leave alone.

If it fires on a passage where the hedge should have gone, the rule is too broad. Add the exception now, while you still remember what it was for.

When you make the same correction twice, write it as a rule someone else could check by looking at the text, then test it on a passage where it shouldn’t apply.

5. Work from a style guide

Apply a guide to a longer piece

Open the community archive report and working style guide. Its four paragraphs give you several kinds of editorial work: a wordy sentence you can simplify, a qualification worth keeping, and a claim about community control that needs the author’s decision. The report and its details are invented.

Read the report first, then try a general edit without the guide.

Edit the report without a style guide
Edit this community archive report for clarity and concision.

Preserve its meaning and voice. Explain the changes.

Supply the complete four-paragraph community archive report first, without the guide. Save the reply, then compare it with a separate review using the guide and the same original text.

Prepared example: “A common description template has made it easier for contributors to compare records.” That simplifies one sentence well. It leaves a separate question about the paragraph: how does the team’s final decision-making authority fit with the claim that contributors retain control?

In a fresh conversation, give it the same original report and the guide rules you want to keep, and ask for a reading against them.

Read against a style guide
This is a short project report for community partners and university colleagues. Read it against the style guide below. Do not rewrite it yet.

Identify up to five places where a change or an author decision would help. For each, quote the passage, name the relevant rule, and explain the issue. Separate wording changes from decisions about meaning.

Also quote one passage the guide gives you a reason to leave alone. Do not invent missing information or assume every sentence needs editing.

[paste the working style guide]
[paste the four-paragraph report]

Paste the four-paragraph community-archive report and the working guide from the editorial session's practice texts. Start with comments so you can choose which edits to accept.

For each comment, check the quotation and the rule. Does the rule apply here? Is the change a matter of wording, or does it decide what the writer means? The prompt also asks for a passage to leave alone, because a guide should protect deliberate choices as well as catch problems.

Each comment now comes with a reason you can argue with. In the prepared example on the slides, the guide protects “may reflect” because the evidence can’t tell several explanations apart, and it raises the question of who controls decisions. Compare that with your first reply.

Then pick one local change you agree with and ask for it.

Propose a small edit
Suggest the smallest edit to this passage that satisfies the style guide:

[paste the passage you chose]

Show the original and proposed wording, name the rule you applied, and explain any change in meaning. If the edit requires a fact or decision I haven't supplied, ask a question instead. Leave the rest of the report unchanged.

After reviewing the comments, choose a specific passage. The full report stays available for context.

Compare the proposal with the original. For example, shortening “the implementation of a common description template” may preserve the point. Replacing “retain control” with “have a voice” changes the claim and needs the writer’s agreement. The prepared examples give you a few comparisons to check once you have made your own.

Ask for a reading against your guide before any edits, with each comment quoting the passage and naming its rule. Then ask for one small change at a time, and for a question whenever a change would need your decision.

Reuse a guide across semesters

The Campus History style-guide exercise uses the actual guide for the student essay collection. It carries editorial decisions across semesters: consistent headings, captions, citations, and light prose cleanup, while preserving each student’s argument and voice. Download its snapshot and try it on one essay you have permission to share.

Reuse a collection's editorial guide
Use the attached Campus History Essay Style Guide to review this essay for publication in the collection. First propose changes, grouped as formatting, light prose edits, and questions for the editor. Name the guide rule behind each proposal.

Preserve the student's authorship, argument, evidence, distinctive phrasing, and justified uncertainty. Normalize presentation without making every essay sound alike. Flag missing citations and changes that would alter a claim. Ask if guide rules conflict.

After I approve the scope, edit a copy and give me a change list. Do not publish or change other essays. Check the rendered page as well as its text before calling the work finished.

Attach the Campus History Essay Style Guide and one essay you have permission to share. Test one essay before applying the guide across a collection.

Prepared illustration: removing “In conclusion,” from “In conclusion, the letters suggest that students may have used the room after hours” follows the guide’s treatment of assignment scaffolding. Keeping “suggest” and “may have” preserves the limit of the evidence. (The sentence is invented, not taken from a student’s essay.)

A saved guide needs checking too. This exercise points out two instructions in the current guide that contradict each other. Look for conflicts like that before you apply a guide across a whole collection, and read the rendered page afterwards, not just the file.

Audit formatting as well as prose

The same approach works on formatting. The CV audit exercise uses a four-page fictional CV in Word that looks finished but has small inconsistencies, such as mixed date ranges and an article listed under two different statuses. Its three prompts audit the CV, check its Word structure, and draw a style guide from its own patterns. The first asks the tool to say whether it read the page, the extracted text, or the Word file itself.

Build a guide from your own writing

The guide in the first exercise belongs to the archive report. For your own work, combine rules from your corrections with a few passages whose style you want to keep. Ask the assistant to describe their patterns from evidence:

Draft your style sheet
Here are three passages I've written that I'm happy with.

Describe my style as a set of rules an editor should follow: typical sentence length and rhythm, how I handle hedging and qualification, how formal I am, what I do with evidence and quotation, and any habits you notice that I may not have noticed.

Write it as instructions to an editor, not as a compliment to me.

It will get part of this wrong. Correcting it is how the style sheet becomes yours rather than its.

Expect it to get some of this wrong. It often reads caution as vagueness, and it may miss the habits you most want named. Saying why its description is wrong is how you find the rules you want.

Then add rules about what not to do, which tend to change the output more than positive ones:

The never-do list
Add a "never do this" section to that style sheet:

- Never remove my hedges. If a claim is qualified, the qualification is doing work.
- Never replace my terminology with more common synonyms.
- Never add transitions like "Moreover" or "Furthermore" to paper over a gap.
- Never use em-dashes to join independent clauses.
- Never compliment the writing before commenting on it.

These often change the output more than positive rules do. Add the habits you most want gone.

Finally, put the standing instruction at the top, so you don’t have to set out your editorial stance again in every new conversation:

The standing instruction
When I bring you my writing, you are my editor, not my ghostwriter.

- Comment; don't rewrite. If I want new prose I'll ask for it explicitly.
- Quote the text you're commenting on. If a point is general advice rather than about this draft, label it as general.
- Say what confused you, not what you would have written instead.
- Don't open with praise. Start with the problem.
- If you're unsure what I'm trying to do, ask before diagnosing.

Then apply my style sheet, below.

Keep this in a text file and paste it at the top of any editing session, or put it in a Claude Project or a custom GPT so it applies every time. It collects the instructions from every step above into one block.

6. Move out of the chat window

This step is optional. Everything above works in a browser tab with nothing set up, but two things you made today won’t outlast the conversation. Your rules scroll out of view, and your draft is scattered across dozens of messages in someone else’s tool. Keep both somewhere more permanent.

Give the rules somewhere to live

The simplest home is a plain text file that you paste at the top of a new conversation. It depends on nothing, and you can move it into a tool later.

If you want the rules applied automatically, use a project, which holds instructions and applies them to every conversation inside it: Projects in Claude, projects or custom GPTs in ChatGPT, Gems in Gemini. Make one per piece of work rather than one for everything, since a book chapter and a grant narrative need different briefs.

What goes in the project
You are my editor for everything in this project. My standing instruction and style sheet are below. Apply them without being reminded.

[paste your standing instruction]
[paste your style sheet]

About this project: I am working on [what], for [audience], by [when].

The current draft is whatever I paste at the start of a conversation — not a version you remember from an earlier one. If you aren't sure you have the current text, ask for it.

When I paste a draft, wait for me to say what I want read. Don't comment on the whole thing unless I ask.

Paste this into the project's instructions field (a Claude Project, a ChatGPT project or custom GPT, a Gemini Gem). You write it once, and it applies to every conversation inside the project.

A project remembers everything you have given it, including the draft you replaced three weeks ago, which is why the instructions above tell it to ask which version to use. As with anything you paste, upload texts you have permission to share.

Keep the draft in a file

Write in a plain text file: .txt, or .md if you want to keep headings. A plain file opens in any program, will outlast whichever assistant you use this year, and makes it obvious which version is current. Treat the conversation as scratch paper and the file as the draft.

Keep the original file to compare against. Ask for a reading of one part, decide which findings to act on, then either edit it yourself or ask for one specific change to a copy. Read what changed before you keep it. A list of changes shows you how much was touched, not just how it was worded.

A pass on one section
Here is my full draft. I am only working on [section / the paragraph beginning "..."] right now. Everything else is context. Don't comment on it.

Read that section against what surrounds it. Where does it repeat something I have already said, contradict something later, or promise something I never deliver?

Quote the specific text. Don't rewrite.

Paste the whole file, then this. The rest of the draft is context: it stops the tool solving a problem the next paragraph already solves.

Some tools can edit your files directly, without copying and pasting. That’s faster, but copying and pasting was also a chance to look at each change. When one instruction can rewrite a whole draft, ask for a list of changes and work somewhere you can undo.

Make the changes visible
Don't give me a revised version of the file. Give me a numbered list of the changes you would make: the exact text as it stands, the exact text you would put in its place, and why.

I will make the edits myself.

The risk with a file is a clean new version with three sentences quietly changed. Ask for the list instead of the file.

Ask for comments instead of changes

Some tools can return a copy of your Word file with comments attached to the passages they’re about, the way a colleague marks up a draft. Nothing in your text changes. You read each comment beside its passage and decide what to do with it.

Ask for comments in the document
Read the attached Word document as [a specialist in the field / a reviewer for a journal / a first-year student].

Add native Word comments anchored to the exact passages you would question. For each comment, say what is wrong, unclear, or unsupported, and what the author could check. Clearly identify any inference.

Preserve the document's text, styles, layout, and formatting. Do not rewrite sentences or add tracked changes. Return a new copy with your comments.

If you cannot return native Word comments, say so and give me a two-column table containing the exact passage and the proposed comment.

Attach a Word file and pick the reader. Your text comes back unchanged, so you can read each comment beside its passage and decide what to do with it.

In this example, a small model running locally (Gemma 4, through Ollama) wrote a study guide to Machiavelli’s political thought, and Claude added 37 comments to it in Word. Here are three of them, with the passages they’re attached to:

Three of Claude's comments, with their passages Claude Fable 5.1 · 2026-09-23

This is a monumental and complex topic. Machiavelli is not just a political thinker; he was a radical paradigm shift. He moved political theory away from the moral ideals of classical philosophy (Plato, Cicero) and into the messy, pragmatic reality of power.

Overstated and partly backwards. Machiavelli’s break is with the moralizing humanist tradition (the Ciceronian ‘mirror for princes’), not with classical antiquity as such. He is saturated in Roman sources — Livy above all, plus Polybius, Sallust, Tacitus, Xenophon, Lucretius — and the preface to Discourses Book I explicitly calls for imitating the ancients. The precise point is that he inverts Cicero’s De officiis rather than abandoning it. ‘Radical paradigm shift’ is filler; the concrete claim is Prince ch. 15’s ‘verità effettuale della cosa’ (the effectual truth of the matter) set against ‘imagined republics and principalities.’ That phrase is the manifesto of the whole project and appears nowhere in this guide.


To build a comprehensive guide, we must move beyond the common—and often inaccurate—caricature of the “Machiavellian villain.” We are analyzing a sophisticated treatise on statecraft and human psychology.

Good framing — but the guide’s own Study Checklist ends by declaring ‘The survival of the state justifies the means,’ which is exactly the caricature this paragraph promises to move beyond. Internal inconsistency; see comment there.


Before studying his ideas, you must understand the world he lived in. Machiavelli was a product of the tumultuous Renaissance Italian city-states (Florence, Milan, Rome).

Rome was the seat of the Papal States, not a city-state in the sense Florence and Milan were. The standard five-power system of the Peace of Lodi (1454) is Florence, Milan, Venice, Naples, and the Papacy. Venice and Naples matter here: Venice is the republic Machiavelli measures Rome against throughout the Discourses, and the Neapolitan succession is what brought Charles VIII over the Alps in 1494.

Copied from the returned Word file, which has 37 comments in all. The study guide was written on September 18, 2026 by Gemma 4 (e4b), a small model run locally through Ollama, and is shared with permission. The prompt for this run wasn't saved; the one above is the workshop's version.

Not every tool or account can return a Word file with comments. If yours can’t, the last line of the prompt asks for a table of passages and comments instead.

Keep the draft and your rules in files you control, and bring them to each new conversation. Ask for a list of changes, or for comments instead of changes, so you read each one before it reaches your draft.

What to take with you

Use these five techniques on a syllabus, report, abstract, CV, or shared collection. Start with one passage or one recurring problem.

  • Brief the task. Name the audience, purpose, and limits. Supply the current text and say what must stay intact.
  • Ask open questions, then anchor. Broad questions bring in advice you didn’t think to ask for. When a comment is about your text, ask which passage it refers to and which rule makes it a problem; a quotation on its own doesn’t prove the criticism.
  • Compare readers. Keep the text fixed and change the reader. Turn the differences into questions, then check them against the text and against real readers.
  • Save your decisions. Turn corrections you keep making into a short, reusable guide. Say what to preserve and when to ask, and rewrite rules that conflict or fire too widely.
  • Work in passes. Diagnose, choose an edit, then read the result. Keep the original, and check both meaning and formatting before you accept a change.