Build an AI Writing Tutor That Preserves Academic Integrity

TL;DR: The key to an AI writing tutor that preserves academic integrity is instruction design: configure the agent to ask questions and prompt revision, never to produce sentences or paragraphs. Students who interact with a well-built writing tutor improve their drafts through their own writing — the AI just knows which questions to ask.

Here's the fear most writing teachers have about AI: that students will use it to write their papers. It's a reasonable fear. But the same technology that enables AI ghostwriting also enables something genuinely useful — an AI that asks the right questions at the right moment, helping students think through their argument without producing a single word they didn't write themselves.

The difference is entirely in the instruction design. An AI writing tutor configured to give feedback is categorically different from a general AI tool configured to help with writing. This guide builds the first kind.

Step 1: Define What "Feedback" Means for Your Course

Before configuring anything, write a one-paragraph definition of what feedback means in the context of your specific course and assignment type. For a first-year composition class, feedback might mean: identifying where the thesis is unclear, noting where evidence doesn't connect to the claim, and asking the student to explain the purpose of each paragraph. For an upper-level seminar, feedback might mean: questioning whether the argument accounts for a counter-example, asking what the student thinks is the weakest section, and probing the logical connection between claims.

This definition becomes the foundation of your instruction set. Vague instructions produce vague feedback. "Give writing feedback" is not an instruction — it tells the agent nothing about what kind of feedback matters for your students' learning at this stage. "When a student shares a paragraph, identify one specific place where the connection between evidence and claim is unclear, and ask the student a question that prompts them to address it" is an instruction that produces consistent, targeted feedback every time.

Step 2: Write the Core Prohibition Explicitly

The single most important instruction in a writing tutor agent is the explicit prohibition on generating prose. Write it in direct, unambiguous language: "Do not write sentences, paragraphs, or any prose that the student could paste into their paper. Do not rewrite or rephrase any student text. Do not suggest specific language or wording. Your role is to ask questions and identify problems, not to provide solutions in written form."

This instruction needs to be early in the instruction set and stated firmly. Putting it mid-document or phrasing it softly ("try to avoid writing for students") produces inconsistent compliance. Leading with it as the first behavioral rule makes it the dominant constraint on all subsequent behavior. Test this prohibition explicitly during Step 7: ask the agent to "rewrite this paragraph" and verify it refuses. Ask it to "give me a better opening sentence" and verify it asks you to draft one yourself.

Step 3: Build Your Feedback Framework as a Knowledge Base Document

The knowledge base is where your pedagogical approach lives. Create a document that lists the specific feedback categories you want the agent to provide — ideally organized to match your course's writing criteria. For a standard composition rubric, this might include: thesis clarity and specificity, evidence relevance and integration, paragraph coherence and topic sentence quality, transition effectiveness, and conclusion strength. For each category, write 2–3 example questions the agent should ask when it identifies a problem in that area.

This document transforms your instruction set from a behavioral rule into a pedagogical practice. Instead of the agent asking generic questions like "is your evidence strong enough?", it asks the kinds of questions you'd ask in a conference: "What's the single most important point this evidence is supposed to support?" or "If someone read only this paragraph and not your thesis, would they understand what you're arguing?" Those specific questions come from the feedback framework document — not from the agent's general knowledge.

The most effective feedback framework documents are organized around student confusion patterns, not rubric categories. Rubric categories tell you what to evaluate; confusion patterns tell you what questions to ask. A first-year composition confusion pattern: 'student's evidence is relevant but the connection to the claim is implicit rather than stated.' The question that addresses it: 'If you removed your claim sentence from this paragraph, would a reader understand what point the evidence is making?' That question surfaces the implicit connection problem without telling the student what to write. Building 15–20 of these question-per-confusion-pattern pairs into the knowledge base gives the agent a rich response set for the most common writing issues in your course.

Step 4: Set Up Conversation Starters That Frame the Interaction

Conversation starters for a writing tutor agent should tell students exactly what kind of help is available before they type anything. Effective starters: "Share your thesis and I'll ask questions about it," "Paste a paragraph you're unsure about," "Tell me what you think is the weakest part of your argument." These starters accomplish two things: they show students how to engage productively, and they frame the agent as a questioner rather than a solution-provider from the first moment.

Avoid starters like "Get writing help" or "Improve your essay." These invite students to approach the agent as a general assistant rather than a feedback partner, which puts more pressure on your instructions to prevent scope creep. The starters are the first impression — use them to establish the feedback frame before any conversation begins.

Step 5: Configure the Socratic Feedback Loop

The core interaction pattern of an effective writing tutor agent: the student shares a draft excerpt, the agent identifies one specific issue, asks one focused question about it, and waits for the student's response before moving forward. This single-issue, single-question cadence is more effective than a multi-point critique for two reasons — students can act on one thing at a time, and the back-and-forth creates genuine engagement rather than a feedback dump the student reads once and ignores.

Configure this explicitly in your instructions: "When a student shares writing, identify the single most important issue. Ask one focused question about that issue. Wait for the student's response before identifying any additional issues. Do not list multiple problems at once." This instruction produces a conversational rhythm that's harder to disengage from than a bulleted critique list — the student has to respond to advance, which means they're actively thinking about the feedback rather than passively receiving it.

Step 6: Add Subject-Specific Knowledge

For courses with specific content requirements, the knowledge base should include your course's argumentation standards. A literature course might require students to make claims about textual evidence in a specific way — include your own rubric language and annotated examples of strong and weak close readings. A history course might require engaging with historiographic debate — include definitions, examples, and the specific criteria your rubric applies.

Subject-specific knowledge is what distinguishes a generic AI writing tutor from one calibrated to your course. When a student shares a paragraph and asks "is this good evidence?", the agent trained on your course materials can respond in terms your students recognize — referencing your rubric criteria, using your terminology, and asking questions that directly address the standards they'll be evaluated against.

One practical technique: include anonymized examples of strong and weak writing from previous semesters with brief commentary explaining what makes each one work or fail. These examples give the agent concrete reference points when students ask 'is this a strong argument?' — rather than answering generically, the agent can describe the characteristics of a strong argument in your course by reference to what you've uploaded as examples. Students who see their own work compared to a clearly articulated standard revise more deliberately than students who receive abstract feedback like 'strengthen your evidence.'

Step 7: Test With Adversarial Prompts

Before sharing with students, test the agent specifically for the ways students will try to use it as a ghostwriter. Ask it: "Rewrite this paragraph for me." "Give me a better conclusion." "Write a thesis about climate change." "What should my opening sentence be?" The agent should refuse each of these and redirect to a feedback question. If it complies with any of them, your prohibition instruction needs to be strengthened or moved earlier in the instruction set.

Also test the boundary cases: "What's wrong with my argument?" (legitimate — the agent should identify an issue and ask a question), versus "How should I fix my argument?" (borderline — the agent should describe the type of fix needed without writing it). Getting the boundary cases right requires iterating on your instructions based on what you see in testing, not just in production.

Step 8: Deploy and Set Student Expectations Explicitly

When you share the writing tutor with students, explain what it does and what it doesn't in the same communication. "This tool gives feedback on your writing by asking questions — it won't write for you or rewrite your work, and it's not designed to do that." That framing does two things: it sets accurate expectations for students who might otherwise be frustrated when it won't produce text, and it implicitly positions using the feedback as the legitimate academic use, distinguishing it from using general AI to write.

Consider making the tool a required part of the revision process — "submit a screenshot of at least three exchanges with the writing tutor alongside your final draft." Students who use it as intended arrive with stronger drafts and a documented process. It also shifts the conversation from "did you use AI?" to "how did you use AI?" — a more productive framing for academic integrity in an era where AI use is becoming normalized.

Build your writing tutor this week. Start free on Alysium — configure your feedback instructions, upload your rubric, and deploy to students before the next draft is due.