1. Workshop Goals and Use Cases

By the end of the workshop, each teacher should leave with a small tool or repeatable workflow related to their own work. The outcome does not need to be a complete system or a large codebase. It may be a process for organizing student data, extracting web content, generating teaching pages, building interactive practice, or packaging a recurring administrative task as a reusable Skill.

Every teacher’s context is different. Different courses, schools, student groups, and administrative needs create different pain points. The purpose of the course is not to provide one standard answer, but to help participants understand the capability boundaries of AI Agents. If a task can normally be done with keyboard, mouse, browser, files, and simple tools, there is a chance an AI Agent can help plan, execute, and document it.

The workshop uses Codex as the core tool, alongside VS Code, GitHub, GitHub Copilot, Playwright MCP, and Ollama. The point is not to memorize tool names, but to understand how they can be combined into a workable process.

3. Natural Language, Vibe Coding, and the Changing Role of Code

The way people interact with computers is changing. In the past, interaction mainly depended on keyboards, mice, command lines, and graphical interfaces. Now natural language is also becoming an interface for operating computers.

Vibe coding is not only about coding. Its core is explaining clearly what you want. The user provides goals, materials, constraints, and feedback; AI helps generate programs, pages, teaching materials, automation flows, or data-processing results.

This does not mean programming knowledge no longer matters. When an outcome will be published, handle sensitive data, connect to external systems, or require long-term maintenance, safety, correctness, and maintainability still need to be checked.

4. ChatGPT: Chat, Work, and Codex

Think of ChatGPT as the entry point, then choose the surface that matches the task. Chat is for quick questions, explanations, and short drafts. ChatGPT Work is for a multi-step deliverable that may use files or approved tools. Codex is for work that needs a codebase, local project files, commands, or developer tools.

These surfaces are complementary, not a claim that ChatGPT only produces text and Codex alone creates files. ChatGPT Work can also use files and approved tools. This workshop uses Codex for exercises that require a local workspace, repeatable scripts, and verifiable outputs.

Whenever a task can access files, tools, or external services, confirm its workspace boundary, permissions, and required approvals before continuing.

5. Data Privacy, Cloud AI, and Local LLMs

Teachers often handle student data, grades, administrative data, and internal school documents. These materials should not be uploaded directly to cloud AI services without proper handling.

Before processing sensitive data, de-identify it first. Replace names, student IDs, class information, seat numbers, and other identifiers with codes, and keep only the fields needed for analysis.

A Local LLM runs on the user’s own computer, for example through Ollama. Its advantage is that data can stay local; its disadvantage is that it requires more hardware resources and may be slower or less capable than large cloud models.

6. Tokens, Permissions, and Security Awareness

Tokens can be understood as the units AI systems use to process text and context. Longer conversations, more files, and more complex tasks consume more tokens. If an AI Agent repeatedly works in the wrong direction, it may quickly spend the available budget.

Permission management is central to AI Agent use. Users should avoid granting full computer access at the beginning. A safer approach is to limit work to a specific project or workspace so the Agent can read and modify only that folder.

Beginners should not leave an Agent running unattended for long periods. Teachers should watch what it plans to do, what it executes, whether it repeats failed attempts, whether it needs additional permissions, and whether the output matches expectations.

7. Computer Environment and Hardware Requirements

This course mainly uses cloud models, so model inference does not need to run on participants’ computers. A Windows laptop or Mac from the past three to five years is usually enough for the course exercises.

AI Agent work usually involves more than one chat window. Practice may involve Codex, a browser, a terminal, VS Code, a local server, and project folders at the same time. Sufficient memory matters more than a high-end GPU for this workshop.

Common classroom issues include network restrictions and slow downloads. If the ChatGPT desktop app, Codex CLI, packages, browser tools, or sample data download slowly, the cause may be school network policy or shared bandwidth rather than the tool itself.

8. Using Codex in ChatGPT

Open ChatGPT and sign in, then select Codex when the task needs a project folder, code, or developer tools. Available features depend on the account plan, workspace settings, and product rollout, so confirm the options visible after sign-in.

When setting up Codex for the first time, choose a clear project or workspace. Codex looks like a chat tool, but it is bound to a folder. This means it can read files, edit files, and use tools within the specified scope.

Basic settings should prioritize safety and understandability. For non-programmers, it is helpful to ask Codex to explain what it is doing in plain language. File access should initially be limited to the workspace.

Beginners should not enable full access at the start. A better path is to build trust with a small practice project and then gradually expand the task scope.

For details, see: ChatGPT (Codex) Workspace and Permission Settings.

9. Windows 11 Vibe Coding Environment Preparation

This section focuses on Windows 11 laptops. The goal is to prepare a usable vibe-coding work environment, not to master every tool immediately.

The supplementary handout now offers three installation methods, chosen based on permissions, network access, and preference: manual installation, running winget commands one by one; installing via an AI Agent Prompt, by switching the ChatGPT desktop app to Work, turning on Full access, and pasting the course-provided prompt so the AI Agent checks for missing tools, installs them in order, and reports back (switch back to Ask for approval once the install finishes); or the course-provided setup-windows.bat one-click script. All three install the same toolset, so if one method fails, try another or fall back to fixing items manually. After setup, confirm that code, git, node, npm, uv, gh, and codex can show version numbers.

For accounts, confirm that ChatGPT / Codex and GitHub can both sign in. Codex CLI signs in through a browser, and GitHub CLI signs in with gh auth login and a one-time code.

Supplement: Windows 11 Vibe Coding and AI Agent Environment Setup

10. Markdown, AGENTS.md, Skills, and Readable Documents

Markdown is an important document format in AI workflows. It is plain text, but it can organize headings, paragraphs, lists, tables, code blocks, and links. It is easy for people to read and structured enough for AI to use.

AGENTS.md can describe project rules: how to start the project, where files belong, how tests run, what output format is expected, and what data should not be changed.

A Skill is a way to preserve a repeatable process. When a task will happen again, such as organizing a roster, generating a fixed handout format, extracting a specific site, or creating similar reports, the workflow can be written as a Skill.

Practice reference: Introductory AGENTS.md Guide: Simple Project Rules for Codex.

11. Plugins, Playwright MCP, and Computer Use

In Codex, Plugins and Skills are different layers. A Plugin is like a capability or integration package. After installation, it may provide new tools, apps, MCP servers, browser control, desktop operation, or other system capabilities.

A Skill is more like a task guide or repeatable workflow. It does not necessarily add new low-level tools; instead, it tells Codex how to handle a specific type of task, what materials to read, which tools to use, and how to verify the result.

In short, Plugins provide what Codex can use; Skills provide how Codex should work on a task.

Playwright MCP provides structured browser control: opening pages, clicking buttons, filling forms, waiting for JavaScript rendering, taking screenshots, and extracting data.

Computer Use is a more general GUI operation capability. It is powerful but harder to govern, so structured tools such as APIs, CLIs, or MCP should be preferred whenever possible.

12. Record & Replay: From Demonstration to Intent-Based Replay

Record & Replay means recording a user’s operation once, then letting an AI Agent repeat the same intent later. It does not simply memorize mouse coordinates; it tries to understand the task goal, important fields, inputs, outputs, decision points, and completion criteria.

Record extracts the workflow from a demonstration. Replay uses that workflow to complete the task again when the user provides a new date, file, or query condition.

This is different from a traditional macro. Macros often remember fixed clicks and keystrokes; if the screen changes, they may fail. Record & Replay is closer to turning a demonstration into task knowledge.

It is also different from traditional web scraping. Web scraping targets HTML, API responses, or DOM structures; Record & Replay is better for workflows people normally perform through a GUI.

Use this only for clear, repeatable, low-risk tasks with human supervision. Do not use it to bypass website restrictions, CAPTCHAs, paywalls, or terms of service.

13. Comparing Three AI Use Scenarios with Questionnaire Data

This section uses the same questionnaire data to compare three AI use scenarios. Recent cloud chat AI tools such as ChatGPT or Claude can read a full .xlsx workbook directly, without requiring users to paste only a few rows or split the workbook into small pieces.

AI Agents such as Codex or Claude Code are suitable for entering a project folder, reading pre-test.xlsx and post-test.xlsx, parsing questionnaire data, checking fields and missing values, calculating reliability and validity clues, and running paired t-tests for matched pre/post respondents.

Local LLMs are used to discuss the value of keeping data on the local machine. Their speed, ability, and context length may be limited, but the privacy benefit is important.

Hands-on guide: Codex Questionnaire Workbook Analysis Practice Guide.

14. Module 3 (Part 1): Download Historical Exam Papers and Answer Keys

This section uses a real, public, no-login data download task: downloading past Four-Year Technology and Two-Year Program Joint College Entrance Examination (統測) exam papers and standard answer keys from the official Technical and Vocational Education Joint College Entrance Examination Center (tcte.edu.tw). The scope stays simple — no login and nothing to bypass — making it a good beginner exercise for AI Agent-assisted source discovery, format confirmation, file download, and verification.

Workflow: ask the AI Agent (ChatGPT Work or Codex) to open the official site, find the historical exam paper page for the Four-Year Technology and Two-Year Program track, choose an academic year, then choose a subject group (for example, Electrical/Electronics) and locate the download links for the exam paper and the standard answer key. Save the downloads to a named folder and confirm the filenames are clear and the files open correctly.

This is also a good teaching moment: the exam paper and the answer key are usually listed as two separate files. Learners should check that the paper and the answer key match the same year and subject group, and note whether the format is PDF or Word. When using Codex, ask it to summarize the downloads into a small checklist (subject, year, exam file name, answer file name, download status) as a verification record.

Official source: Technical and Vocational Education Joint College Entrance Examination Center — Four-Year Technology exam downloads

Hands-on guide: Codex Exam Archive Download and Markdown Question Bank Practice Guide. This guide walks learners through using Codex to download the 109–115 exam papers and answer keys for the "04 Electrical and Electronics Group — Information/Electronics Track," then convert the Word exam files into a uniformly structured Markdown question bank.

15. Module 3 (Part 2): Question Cropping and JSON Answer Key

This section continues from the exam papers and answer keys downloaded in Module 3 (Part 1) and practices a different kind of data-organizing task: asking the AI Agent to crop the exam PDF into one image per question and build a JSON-format answer key by cross-referencing the standard answer PDF. Unlike "follow the URL and click," this time Codex has to judge for itself where each question starts and ends on the page — a task that depends more on the model's judgment than a fixed procedure.

Workflow: ask the AI Agent (Codex is recommended) to convert the exam PDF into high-resolution images, scan each page to locate where every question starts and ends (including the stem, all options, and any accompanying figures or tables, up to the next question number), crop and name each image using the specified filename rule (e.g., <year>Q<number>.png), then cross-reference the standard answer PDF using the same filenames as keys to assemble an answer_key.json.

This is also a good teaching case: unlike Module 3 (Part 1)'s "download from a fixed URL," cropping requires the model to judge page layout, questions spanning two pages, and two-column formats — a good opportunity to compare precise, detailed, and simple prompt styles, and to ask whether hardcoding details still holds up when a task requires "judging by eye" rather than "following a fixed path."

Hands-on guide: Codex Exam Question Cropping and Answer Key Practice Guide. This guide continues from the "04 Electrical and Electronics Group — Information/Electronics Track, Specialized Subject (1)" exam papers for academic years 111–115 downloaded in Part 1, walking learners through cropping the PDF into one image per question and building the corresponding JSON answer key.

15. From One-Time Operation to Repeatable Workflow

A successful one-time operation is only the beginning. The real value is turning that success into a process that can be run again in the future.

A repeatable workflow should describe the task goal, input conditions, steps, output location, and verification method. For example, a CSV download workflow should explain how to confirm the server is reachable, which URL to open, which course codes are supported, where to save the file, and how to verify the result.

This turns an AI Agent from an on-the-spot chat helper into an accumulative work system. Each completed workflow can become part of the project and make the next task faster and more stable.

16. Capstone Preview: Formative Assessment and AI-Assisted Teaching

The workshop series will end with an AI formative-assessment system as the capstone application. This example is not just an online quiz. It presents a complete teaching-tool scenario: the teacher side tracks class progress, the student side completes short quizzes and reviews responses, and AI helps turn wrong answers and teaching materials into follow-up learning directions.

The point of formative assessment is not to make tests harder. It is to answer three classroom questions quickly: whether students understood the concept, which students or question types need more support, and what students should review next after answering.

The example system includes a teacher dashboard and a student quiz interface. The teacher sees class-level status such as quiz progress, completion, accuracy, and submission state. The student side supports login, question-by-question answering, submission, score summary, and response review.

AI is used as learning scaffolding, not as an answer machine. After students submit their work, their responses, question content, and material summary can be given to ChatGPT, Gemini, or another tool to analyze possible misconceptions, recommend review sections, and suggest next exercises.

Reference: AI Agent Teaching Applications Workshop Capstone Preview: Design Intent for a Formative Assessment System.

17. Q&A and Common Troubleshooting

• Workshop focus: environment setup, project operation, data analysis, web extraction, and repeatable workflows.
• Installation issues: permissions, winget, App Installer, PATH, and network download speed.
• Account issues: ChatGPT / Codex login, GitHub login, quota, and education plans.
• Project issues: wrong folder, missing materials, and unclear output location.
• Questionnaire analysis issues: field interpretation, pre/post pairing, de-identification, and result verification.
• Exam download issues: finding the right year or subject page, matching the exam paper to its answer key, PDF vs. Word format, and clear download locations and filenames.
• Question cropping and answer key issues: whether each crop captures the full stem and options, whether the image count matches the question count, and whether the JSON answer key's keys line up one-to-one with the image filenames.
• Repeatable workflow issues: when to turn work into a document, script, or Skill.