Elective · How I Use AI: Building This Course
Elective · How I Use AI: Building This Course
One real project, three kinds of work: teaching, admin, and research
1 hour, live demo | Date to be announced by school email | No laptop needed; bring questions
The Example: This Website
The site you are reading was built and is maintained with AI. Since June 2026, the pages, slides, PowerPoint files, demo datasets, and the AI Environmental Footprint Explorer have been drafted, built, and checked with Claude, mostly through Claude Code (an agentic tool that works in a folder of files) and the Claude desktop app. Every change is reviewed before it is published, and the full history is public on GitHub.
It is a useful example because one project covers the three kinds of work most of us do: teaching (building course materials), admin (keeping a lot of pages accurate when something changes), and research (finding, checking, and citing evidence).
What You’ll Leave With
- A concrete picture of what an agentic AI tool does well, and where it needs you
- Three patterns you can reuse with any tool: keep context in files, make the AI check its own work, and record what changed and why
- A clear sense of which work belongs in Drexel tools (ChatGPT Edu, Copilot Chat) and which can go in other tools
Run of Show (planned)
| Time | Segment |
|---|---|
| 0:00–0:08 | The setup: tools, folder, and the data boundary |
| 0:08–0:23 | Teaching: rebuilding Core 1 for first-time users (live) |
| 0:23–0:38 | Admin: one fact changes, ten pages need fixing (live) |
| 0:38–0:50 | Research: the Footprint Explorer and its record of corrections |
| 0:50–1:00 | Questions, and how to start with your own project |
The Setup
Why non-Drexel tools are fine here. Everything on this site is public: no student records, no personnel information, no unpublished research. That makes it low-risk information, so tools that aren’t Drexel-supported, like Claude Code, are allowed. The same tools would not be fine for a folder with rosters, grades, or study data. See Tools & Access.
How the work is organized.
- One folder. The site is plain-text files in one folder. The AI reads and edits the files directly, so there is no copying and pasting.
- Instructions in a file. A short project file tells the AI what the course is and who it’s for, so each session starts with that context.
- Handoff notes. At the end of a work session, the AI writes a short note: what was published, what’s waiting for review, and what’s still undecided. The next session starts by reading it. These notes stay on my computer.
- A build and a check. After each change, the site is rebuilt and every link is checked automatically before anything is published.
Example 1 — Teaching: Rebuilding Core 1 for First-Time Users
The task. Core 1 assumed too much for people who had never opened ChatGPT. It needed to be simpler, and the page, the slides, the PowerPoint, and the instructor guide all had to match.
| The AI did | I did |
|---|---|
| Rewrote the page in plainer language, with a glossary and a step-by-step guided tour | Decided what a first-time user needs, and what to cut |
| Rebuilt the HTML slides, and wrote a small script that generates the PowerPoint from the same content, so the two can’t drift apart | Picked the six demo tasks, and ran the live demo myself before the session |
| Made fictional demo data: a crash dataset for a fictional city and a class roster with IDs instead of names | Set the rules for the student-groups example |
| Checked that the example student groups on the slide really meet all three grouping rules | Approved every slide before it went out |
What got caught.
- A citation that looked right. Asked for studies with DOIs, a model returned one DOI that led to a paper on motorcycle helmets. That mistake became a slide in Core 1.
- Steps written from memory. The guided-tour steps for ChatGPT Edu were drafted without seeing the screen. They had to be checked against the real interface before the session.
Example 2 — Admin: One Fact Changes, Ten Pages Need Fixing
The task. The site told people to use Copilot inside Word and Excel. Drexel doesn’t have that license. Every page that said so needed fixing, and Core 2 needed rebuilding around Copilot Chat. The same week, three topics moved into a new Everyday Tools track.
| The AI did | I did |
|---|---|
| Searched every page for anything that implied the license, and listed them | Supplied the facts it couldn’t know: which licenses Drexel has, that faculty can create Canvas tokens, which menu names were right |
| Rewrote Core 2, its slides, PowerPoint, instructor guide, and the text inside the demo files | Decided the new track structure and how basic Core 2 should be |
| Moved three topics to the new track, then fixed topic numbers, counts, and back-links on every page that mentioned them | Reviewed the changes before publishing |
| Rebuilt the site, checked every link (none broken), and wrote a plain description of each change |
What got caught.
- Numbers that were already wrong. The home page said “36 topics across five tracks,” which was already out of date before this change. Finding every mention of something is where an agent beats a person.
- A rule with an edge case. In the Core 2 spreadsheet demo, one row costs exactly $12.00 per person. A formula using “more than” skips it; “at least” catches it. That became a note for instructors.
Example 3 — Research: The AI Environmental Footprint Explorer
The task. After news reports on data-center water use in July 2026, faculty and students asked how much energy and water AI really uses. The Footprint Explorer answers that with ranges, sources, and the disagreements left visible.
| The AI did | I did |
|---|---|
| Research synthesis across agency reports, company disclosures, and journal articles | Traced every figure to a primary source before it went in |
| Wrote the interactive tool, plus R and Stata starter scripts that use a public dataset | Decided which numbers to trust, which to leave out, and how to describe the disagreements |
| Re-checked the figures against new sources every few months | Approved each refresh |
What got caught. The project keeps a public changelog that records, for every changed figure, the old value, the new value, the source that forced the change, and why the old one was wrong. A few entries:
- A widely cited figure was superseded. The often-quoted ~3 Wh per chatbot query came from a 2023 estimate. A 2026 study measured a median of about 0.3 Wh in production, so the high end was lowered.
- A true statement stopped being true. “There is no authoritative estimate of AI’s share of data-center electricity” was correct when written, and wrong after an April 2026 agency report.
- A check that couldn’t be run. The dataset’s website was unreachable when the scripts were written, so the column names were never verified. The scripts print the columns first, and the changelog says so.
Patterns You Can Reuse
These work with ChatGPT Edu and Copilot Chat as well as with agentic tools.
- Keep context in files, not in your head. A short “about this project” note and an end-of-session handoff note save you from re-explaining every time. In ChatGPT Edu, a shared Project does the same job (Core 3).
- Make the AI check its own work, then check the check. Ask for a link check, a count, a rule check, or a list of every place something appears. Then spot-check one or two yourself.
- Give it the facts it can’t know. Licenses, local policy, what the screen really looks like, what your students need. Most corrections on this site came from facts only a person had.
- Record what changed and why. Old value, new value, source, reason. It’s the difference between “updated” and something you can defend later.
Where This Is Heading: AI That Reviews AI
For software, Anthropic’s Code Review runs several AI reviewers on each proposed change. A separate step then checks every finding against the code and drops false alarms, ranks what’s left by severity, and posts it for a person. It never approves a change on its own. The rules for what to look for live in a written file in the project, so every review uses the same checklist.
This site already does the mechanical half: every change is rebuilt and every link is checked before it’s published. The next step is a review pass on the content itself:
- A second reader that didn’t write it. A separate agent, or a fresh chat, reads each change.
- A written checklist. Facts match their sources; the data rules match Tools & Access; plain language; accessible tables and images.
- Evidence for every finding. Each problem quotes the passage and the source it conflicts with, so it can be checked in seconds.
- The author decides. Every finding is read and accepted or rejected by a person.
Code has a head start because it can be run and tested. Documents can’t, which is why the “quote the evidence” step matters so much. Core 4 has a version anyone can try with ChatGPT Edu or Copilot Chat.
Try It Yourself
With ChatGPT Edu or Copilot Chat (everyone): at the end of a work session on a long project, paste your notes and ask:
Write a short handoff note for the next time I work on this: what I finished, what’s waiting on someone else, and what I still need to decide. Use three headings and bullet points.
Paste that note at the start of your next chat.
With Claude Code, Codex, or the Claude desktop app (if you use them): point the tool at a folder of public material, such as your course website or a public report, and ask:
Find every place these files mention [a fact that changed]. List each one with the file name and the sentence. Don’t edit anything yet.
Review the list, then ask it to make the edits and show you what changed. Keep student data, personnel information, and unpublished research out of the folder.
Give Feedback on This Workshop
Your feedback directly shapes how these workshops are revised and what gets scheduled next.
In the first question, select the session you attended.