PROMPT-EDU: Fine-Tuning Command Script for ChatGPT A Structured Prompt Sequence for Optimizing ChatGPT Fine-Tuning in Educational Contexts

Denise da Vinha Ricieri (first author), Adriana M. G. de Farias, Fabiano Rodrigues de Souza, Raphaela V. G. Barreto · SCIAS Educação, Comunicação e Tecnologia — v.6, n.1, p.107–138, jan./jun. 2024 · DOI: 10.36704/sciaseducomtec.v6i1.8374 · Peer-reviewed journal article · e-ISSN: 2674-905X

Research Question

Can a structured, sequenced set of prompts — the Prompt-EDU Script — reliably optimize the fine-tuning (FT) of ChatGPT for educational contexts, including for novice users with accounts that have minimal prior FT history?

Methodology

Design: Qualitative descriptive experimental study tested across multiple ChatGPT versions in Brazil and the USA.
Prompt types used: In-Context Learning (ICL), Chain of Thought (CoT), and Demo prompts combined in a structured sequence.
Three-profile Prompt-EDU Script: Context prompt → Demo-CoT prompts → Educational prompt (using Bloom's Taxonomy verbs).
Test scenario: Simulated dialogue between Paulo Freire and Mark Zuckerberg on AI and Education — same scenario as CoBICET 2023 paper, now used to validate the script's FT performance.
Bias addressed: Non-native language processing bias in GPT (double translation pipeline: input language → EN → output language). Bloom's Taxonomy verbs used to override ambiguous verb interpretation.
Accounts tested: IAR1-BRA (Brazil, GPT-3.5) · IAR2-BRA (Brazil, GPT-4) · IAT-EUA (USA, GPT-4) — beginner FT accounts.

Key Findings

Prompt-EDU Script succeeded on all established analytical markers across all tested ChatGPT accounts and versions.
Performance matched deep machine learning benchmarks described in prior literature for structured prompt sequences.
A single Demo prompt + well-structured CoT prompt outperformed multi-demo ICL — replicating Chen et al. (2023) findings.
Training bias detected and corrected: Brazil account (IAR1-BRA) initially used GPT-3.5 mid-FT — bias isolated and script adapted.
Bloom's Taxonomy verbs effectively controlled AI interpretation of educational verbs across the non-native language (Portuguese) barrier.
Script proven reliable for novice users optimizing FT for teaching-learning themes in Portuguese-speaking contexts.

The Prompt-EDU Script (3 Profiles)

Profile 1 — Context Prompt: Establishes the educational domain, the task, and the AI persona as an educational assistant. Sets the FT direction.

Profile 2 — Demo-CoT Prompt: Provides a Chain of Thought demonstration that trains the model to reason in educational terms before producing output.

Profile 3 — Educational Prompt: Uses Bloom's Taxonomy action verbs to assign a specific cognitive-level task, overriding language ambiguity and producing educationally valid output.

Tested: ChatGPT 3.5 & 4 · Brazil & USA accounts · Beginner FT level

Technical Concepts & Skills

AI & Machine Learning: LLM ChatGPT Fine-Tuning (FT) Deep Learning Transformer Architecture NLP Neural Networks Generalization
Prompt Engineering: In-Context Learning (ICL) Chain of Thought (CoT) Demo Prompts Prompt-EDU Framework Educational Prompts Prompt Sequencing
Education & Pedagogy: Bloom's Taxonomy Instructional Design AI in Education Non-Native Language Bias Teacher Training Systematization

Why This Matters for Educators

Most educators using ChatGPT for lesson planning, content creation, or student feedback are working with accounts that have minimal fine-tuning — and they are writing prompts in non-English languages, which introduces a hidden double-translation bias. The Prompt-EDU Script is the first peer-reviewed, tested methodology specifically designed to overcome both problems: it gives educators a reliable, systematic way to train ChatGPT to behave as a genuine educational assistant, using the international language of pedagogy (Bloom's Taxonomy) to bridge the language gap.

Selected References

Chen et al. (2023) — Many-shot vs. single Demo ICL performance · Wei et al. (2022) — Chain of Thought prompting · Liu et al. (2023) — Pre-train, prompt, and predict (ACM Comp. Surv.) · Wang et al. (2023) — LLMs as implicit topic models · Vaswani et al. (2017) — Attention is all you need · Liang et al. (2023) — GPT detectors biased against non-native writers · Armstrong (2010) — Bloom's Taxonomy revised · Kasneci et al. (2023) — ChatGPT for good? (Learning & Individual Differences) · OpenAI Inc. (2023) — GPT-4 Technical Report.

Read the Full Paper

ResearchGate (open access): researchgate.net/publication/382859526
Journal: SCIAS Educação, Comunicação e Tecnologia — v.6, n.1, 2024 — DOI: 10.36704/sciaseducomtec.v6i1.8374

Author Bio

Fabiano Rodrigues de Souza — PhD in Biotechnology · Harvard Graduate School of Education · Dual MBAs · Kirkpatrick Certified Professional · 2× Braskem Award Winner · Editorial Board Member since 2006. Researcher at the intersection of AI, neuroscience, and multilingual education.

See also — Paper 1: Simulating Dialogues and Characters in ChatGPT-4 (CoBICET 2023) — multilingual GPT-4 evaluation: English vs. Portuguese output quality, double-translation bias, NLP bias taxonomy.
See also — Paper 2: EMO-AI Teaching Loop (UNICAMP X Inovações Curriculares 2025) — neurodidactic microlearning via WhatsApp & GenAI: 21% enthusiasm, 19% curiosity, correlation = 1.0.