A three-part assessment: a Moodle quiz on writing + AI critical use, an agentic AI revision UI, and a teacher review with plain-text reports. Built to answer the AY 2025-26 evaluation feedback.
A quick look back at the original AI-edit module and what last year's data told us, before the plan for this year.
A Nuxt-based AI-assisted Review tool with three modes (briefing / training / assessment), student login, and manual rubric + API-key setup. Students revised their point-of-view essays with an AI and submitted their chat history.
Caveat: both the engagement and essay scores were AI-rated, and the result is correlational rather than causal β treat it as promising and worth re-checking against human-rated essays.
Presented at AI3L 2026 (Japan) β βStructured AI-Guided Essay Revision with Dual-Rubric Assessmentβ (Wang & Deng).
Last year worked: students improved and the βcritical review of AI suggestionsβ idea held up. But the evaluation also told us the module was fragmented (three modes, login, manual setup), the rubrics were vague, and the revision was unfocused. This year's design targets exactly those three weaknesses.
Everything is submitted through Moodle: a compulsory quiz, the AI revision activity, and a Moodle assignment that collects the work. Each stage feeds the next.
Compulsory quiz with MC + short objective questions (understanding + API-key readiness).
Students paste their essay, pick one writing aspect, and get guided revision from an AI tutor that never writes for them β it coaches with questions and requires critical decisions.
Students submit their revision work β revised essay, rubric-mapped report, and chat history β to a Moodle assignment in the main course room. Teachers download the cohort's submissions and review per section.
A compulsory quiz of mostly multiple-choice and short objective questions: it checks understanding of the writing rubric and AI critical use, and confirms each student has their HKBU GenAI API key ready. The revision work itself is submitted via the Part 2 Moodle assignment.
Multiple choice & short objective
Checks understanding of the writing rubric and how to use AI critically β e.g. βWhich is a strong thesis?β / βWhen should you reject an AI suggestion?β
API-key confirmation
A readiness gate, not a knowledge test β confirms each student has their own HKBU GenAI API key ready before they start revising.
1 Β· Compulsory (5%) β every student completes it before the revision activity.
2 Β· Auto-graded β MC and short objective questions are marked automatically, no manual marking.
3 Β· Checks readiness β confirms they understand the rubric and have their own HKBU GenAI API key.
4 Β· Leads into the revision β once done, students move to the revision UI, then submit via the assignment.
A clear minimum keeps marking fair and the task focused:
Minimum 2 rounds β two AI suggestions, each followed by a decision (ACCEPT / REJECT / ADAPT) with a one-sentence reason, within the ONE chosen aspect. More rounds are welcome and strengthen the report, but 2 is the floor.
The activity students do before submitting. Students work with an AI tutor that guides them through targeted revision β never writing for them.
Students already completed the Moodle quiz on the writing rubric + critical AI use β so the tutor goes straight to revising.
A chatbot responds to prompts. An agent has a bounded job, grounded knowledge, connected tools, a designed workflow, guardrails, and an evaluation loop. The tutor drives the student through a deterministic pipeline β it doesn't just chat.
Hosted on Railway. Bring your own HKBU GenAI key (BYOK), or enter the test access code test2026 (staff only) to try it with our key. Anonymous β no login needed.
After revising with an AI, each student submits their work to a single Moodle assignment in the main course room (all sections together).
1 Β· Revised essay
The final version, focused on their ONE chosen aspect (Content / Organisation / Vocabulary / Grammar).
2 Β· Rubric-mapped report
Generated by the in-house tutor (or written following the prompt pack's REPORT step) β mapped to criteria A / B / C with quoted evidence.
3 Β· Chat history (appendix)
The full studentβAI transcript, so markers can verify the report against the conversation.
One submission per student: a single file upload (PDF) or Moodle online text with the three parts pasted in.
Then what happens
Download the submissions β run them through a programme that parses each report β send each teacher their own sections' preliminary reports.
We can embed the tutor directly here. GitHub Pages serves this page; the tutor's backend runs on Railway and is shown inside an iFrame. Try it below β no install needed.
Paste this into a Moodle Page, a course website, or any HTML editor. The UI runs on Railway; the iFrame is just a window onto it.
GitHub Pages is a static file host β it can serve HTML, CSS, and JavaScript, but it cannot run a Node.js server or safely hold the HKBU GenAI API key. So the "brain" of the tutor runs on Railway, and any page (this one included) just displays it through an iFrame.
The iFrame is only a window β the agentic work happens on the other side.
The revision activity isn't locked to our UI. Students can do it in any AI chat platform β including the HKBU GenAI chatbot UI β using a portable prompt pack.
Students paste the portable prompt pack into the HKBU GenAI chat interface. Same workflow, same rubric, no extra setup.
Students paste the prompt pack into DeepSeek or Doubao. The AI follows the same three-step revision + critical-pickup workflow.
Any OpenAI-compatible chat UI works. The prompt pack is platform-agnostic β the pedagogy travels with the text.
Paste this whole block into any AI chat, then replace the two placeholders. It encodes the same job, workflow, guardrails, and rubric as the in-house tutor.
Full version (incl. a stepped 4-prompt variant and the two question sets) β generic-chatbot-tutor-prompts.md
A prompt reproduces the pedagogy, not the enforcement or tooling. Roughly 60β70% of the learning experience transfers for a motivated student.
| Block | Agentic platform | Prompt-only chatbot |
|---|---|---|
| Job | β enforced | β οΈ soft β model may drift |
| Knowledge | β injected | β οΈ only what student pastes |
| Tools (report / submit / diff) | β automatic | β manual copy-paste |
| Workflow | β step-gated | β οΈ prompted, can be skipped |
| Guardrails (no ghostwriting) | β hard | β οΈ soft β easily overridden |
| Evaluation loop | β link β teacher β data | β no automatic loop |
Recommendation: ship both. The platform is the recommended path (convenience + integrity + teacher review); the prompt pack is the equity fallback and an AI-literacy lesson (the prompt is the tutor's program).
Submissions land in the main course room assignment. A programme parses them and produces a preliminary report per section, which each teacher reviews in Moodle.
Pull the whole cohort's submissions from Moodle in one go.
Run them through a programme that maps each to the rubric.
Send each teacher their own sections' preliminary reports.
π Rubric-mapped report
A. Conversation depth Β· B. Critical pickup (PRIMARY) Β· C. Refining process β with quoted evidence from the chat
π¬ Chat transcript
Full conversation between student and AI tutor
π₯ Plain-text download
One-click .txt export for records or Moodle feedback
Weights A = 3 Β· B = 5 (PRIMARY) Β· C = 3 β total 11, scaled to the 5% assignment. Full version in the Google Doc.
| Criterion | Weight | What it scores |
|---|---|---|
| A. Negotiating with the AI | 3 | How actively the student engages the AI as a thinking partner β follow-ups, pushback, alternatives |
| B. Critical Review of AI Suggestions β PRIMARY | 5 | Whether the student decides (ACCEPT / REJECT / ADAPT) and justifies each suggestion, instead of accepting blindly |
| C. Refining Process | 3 | Visible improvement in the ONE targeted writing aspect, shown as a before/after pair |
| Band | Descriptor |
|---|---|
| 1 β Limited | Pastes the essay and accepts the first response; asks no questions; no real conversation. |
| 2 β Basic | One or two brief exchanges; at most one simple question; mostly follows the AI's lead. |
| 3 β Developing | A few relevant follow-up questions; occasionally asks "why?" or for an example; some back-and-forth. |
| 4 β Proficient | Sustained back-and-forth; asks "why?" and "what if Iβ¦?"; pushes back at least once; requests an alternative. |
| 5 β Excellent | Drives a genuine negotiation β challenges suggestions with their own reasoning, proposes alternatives, and builds on earlier turns. |
| Band | Descriptor |
|---|---|
| 1 β Limited | Accepts every suggestion without comment; no reasons; no sign of evaluation. |
| 2 β Basic | Accepts most suggestions with only vague or generic reasons ("sounds good"). |
| 3 β Developing | Makes decisions (ACCEPT / REJECT / ADAPT) on some suggestions, with brief reasons that may be generic. |
| 4 β Proficient | Decides on most suggestions with clear, text-specific reasons; justifies at least one rejection or adaptation. |
| 5 β Excellent | Decides on every suggestion with a thoughtful, evidence-based reason; rejects or adapts where appropriate and explains how that preserves their own meaning and voice. |
| Band | Descriptor |
|---|---|
| 1 β Limited | No genuine revision; the essay is unchanged, or the changes are copied straight from the AI. |
| 2 β Basic | Minimal, surface changes (a word or two); improvement is unclear. |
| 3 β Developing | Some revision in the chosen aspect; improvement is partial or inconsistent. |
| 4 β Proficient | Clear improvement in the chosen aspect, shown by a before/after pair; the changes are the student's own. |
| 5 β Excellent | Strong, meaningful improvement in the chosen aspect; the before/after shows real gains and the student can explain the change. |
The core skill of the AI-assisted Review is critical thinking about AI advice β not the AI doing the work. The tutor requires a decision (ACCEPT / REJECT / ADAPT) with a reason after every suggestion, and the report quotes those decisions as evidence. This is the anti-ghostwriting guarantee.
Students focus on one writing aspect at a time (Content & Ideas, Organisation, Vocabulary, or Grammar). This makes the improvement visible and answers the evaluation feedback about insufficient scaffolding.
The activity works with any AI chatbot of the student's choice β HKBU GenAI, ChatGPT, Claude, Gemini, DeepSeek, Doubao, and others. The rubric is identical; only the tool changes. Full version in the Google Doc.
The rubric is platform-neutral β A / B / C apply no matter which tool the student used.
Assess β run the Gen AI rater prompt over each submission, then review and adjust.
Look for β B: decisions + reasons Β· A: follow-ups & pushback Β· C: a genuine before/after.
Anti-ghostwriting checks
Verify the changes are the student's own; the reflection should explain them. Watch for AI rewrites accepted wholesale.
This plan is ready for a small-group pilot. Share your thoughts.
Questions, suggestions, or concerns about the AI-assisted Review plan? We review all feedback.
Open GitHub Discussions βOr email the course team directly.