Take-home projects.

For: Anyone who runs meetings

You need: ChatGPT Edu, plus one cloud-recorded Zoom meeting with “Create audio transcript” enabled (turn it on today in the Zoom web portal under Settings → Recording; it only applies to future recordings).

Steps

1. Enable the audio transcript in cloud recording settings
2. Record your next routine meeting to the cloud, with everyone's knowledge; participants see Zoom's accept-or-leave prompt
3. In the web portal, open Recordings and download the .vtt transcript file
4. Skim it once and delete any passage you wouldn't put in minutes
5. Upload to ChatGPT with the starting prompt
6. Check the action items against your own memory
7. Read the follow-up email draft, fix anything off, and send it from Outlook

Starting prompt

This is a meeting transcript in VTT format with timestamps. Produce three sections: (1) a five-sentence summary; (2) action items as a table with owner, task, and deadline. Use the role or name exactly as spoken, and write “unassigned” if no one took it; (3) decisions made, one line each, with the timestamp where the decision happened; (4) a follow-up email to attendees, ready to send: a two-line recap, the action-items table, and the decisions. Do not invent owners or deadlines that were not stated.

Caution: Recordings capture voices. Record only with the consent prompt shown, and never upload a transcript containing student-record content or anything Confidential. Edit the .vtt before upload if in doubt.

Working looks like: The timestamps in the decisions log let you jump to the exact moment in the recording and confirm each one is real, and the follow-up email needs one read before it goes out.

For: Faculty and research staff

You need: ChatGPT Edu with scheduled tasks, and a topic you genuinely want to follow.

Steps

1. Open the Scheduled page in the ChatGPT sidebar
2. Create a weekly task, Monday 7am
3. Paste the starting prompt with your topic
4. Run it once immediately to test
5. Spot-check two of the items it cites: do they exist?
6. Tighten the prompt with whatever it got wrong (too broad, wrong venues)
7. Let it run for three weeks before judging it

Starting prompt

Search the web for material published in the last 7 days on [topic]. Return the 5 most substantive items: title, source, one-paragraph summary, and why it matters for [my role]. Prefer journals, university announcements, and major outlets over blogs. If fewer than 5 substantive items exist, return fewer; do not pad.

Caution: None on data, since this touches only public information, which makes it the safest scheduled task in the set. Do verify links; fabricated citations are the known failure mode.

Working looks like: Three consecutive Mondays where every link resolves and at least one item was genuinely new to you.

For: Faculty

You need: ChatGPT Edu's GPT builder and your syllabus as a PDF or DOCX.

Steps

1. Strip anything from the syllabus you don't want repeated verbatim (personal cell number, home office location)
2. Open GPTs → Create, and work in the Configure tab directly
3. Name it “[Course number] Assistant”
4. Paste the instructions text below
5. Upload the syllabus under Knowledge
6. Turn OFF web browsing and image generation, so it stays grounded in the file only
7. Test with ten real student questions, including two it should refuse (“what will be on the exam?”, “can you do question 3 for me?”)
8. Before sharing with students, follow the approval process: sharing a custom GPT requires a submitted Clemson request first

Starting prompt

Instructions text: You answer questions about [COURSE] using only the uploaded syllabus. Always cite the syllabus section. If the answer is not in the syllabus, say so and direct the student to email the instructor. Never guess about dates, grading, or policies. Do not complete assignments, write drafts, or discuss exam content. Tone: friendly, brief.

Caution: The GPT is only as private as its knowledge files: never upload a gradebook, roster, or accommodation letters. Sharing beyond yourself requires institutional approval first.

Working looks like: It answers “when is the midterm?” with the right date and a section citation, and refuses “write my reflection paper.”

For: Faculty and research staff

You need: ChatGPT Edu Projects, and 5 to 20 PDFs you have the right to upload: your own writing, university reports, OER, public documents.

Steps

1. Create a Project named for the effort (“Accreditation self-study”)
2. Upload the PDFs as project files
3. Set the project instructions: answer from the uploaded files, always name the file and page, say so when the files don't cover it
4. Ask three questions you already know the answers to; that's your calibration
5. Ask the real questions
6. For each answer you'll act on, open the cited file and confirm
7. Share the Project with a colleague working the same material if useful

Starting prompt

Across all uploaded files, what do we say about [X]? Give me every relevant passage with file name and location, then a two-paragraph synthesis, then contradictions between documents if any.

Caution: Copyright is the constraint here more than privacy: publisher journal PDFs generally shouldn't be uploaded to any AI tool. Your own manuscripts, reports, and OER are fine.

Working looks like: The three calibration questions come back correct with accurate file-and-page citations.

For: Staff with heavy inboxes

You need: ChatGPT Edu with the Outlook connector enabled. Connectors are admin-managed and need a one-time CheckIT assessment request; submit it naming the use (inbox triage and reply drafting) and this whole project unlocks.

Steps

1. Submit the CheckIT request for the Outlook connector, one time
2. Once your workspace admin enables it, connect Outlook in ChatGPT settings
3. Create a Project called “Inbox” with standing instructions: triage unread mail, draft a reply for everything quick, never send
4. Upload two or three of your sent emails as the voice sample
5. Each morning, run the standing prompt and read the drafts
6. Send each approved draft from Outlook; the AI never sends anything itself

Starting prompt

Read my unread Outlook mail from the last 24 hours. Sort into: answer-in-two-minutes, needs-a-real-block, delegate, ignore. For every answer-in-two-minutes item, draft the reply in my voice, matching the uploaded samples, three sentences max. List the needs-a-real-block items with a one-line reason each. Draft only; do not send anything.

Caution: The connector reads the whole mailbox, which is exactly why it goes through review first. Student situations and personnel matters still get your eyes before any draft goes out.

Working looks like: You open the morning run to a triaged list and drafts you can send with one edit, and nothing has ever gone out without your approval.

For: Staff and faculty on committees

You need: ChatGPT Edu and a meeting packet you're allowed to process.

Steps

1. Create a Project per committee
2. Set instructions: what's being decided, what's information-only, what each attachment says in three sentences, what a diligent member should ask
3. Before each meeting, upload the packet PDFs and run the standing prompt
4. Read the two attachments flagged as decision-relevant yourself
5. Bring the generated questions list

Starting prompt

Here is the packet for [date]'s meeting. Give me: (1) items requiring a vote, with a plain-English statement of what a yes means; (2) three-sentence summaries per attachment; (3) five questions I should be ready to ask, ordered by importance; (4) anything in the fine print that contradicts the cover memo.

Caution: Packets marked confidential (personnel, legal, donor) don't go in. The contradiction check is the highest-value and least-reliable output, so verify before raising it in the meeting.

Working looks like: You walk in having read 20% of the packet and ask the best question in the room.

For: Staff running events

You need: ChatGPT Edu and one source document: event description, speaker bio, date, venue, registration link.

Steps

1. Write the single source of truth: title, speaker, date, time, venue, link, and three facts about why it matters
2. One prompt generates the full kit
3. Check every date, name, and link against your source, because copy errors here are public
4. Adjust one output's voice, then say “match this voice for the rest”
5. Save the working prompt in a Project called “Events” for reuse

Starting prompt

Here are the facts for an event; treat them as the only source of truth and change nothing: [paste]. Produce: (1) 150-word email announcement; (2) 50-word newsletter blurb; (3) three social posts of increasing urgency (3 weeks out, 1 week, day before); (4) a 30-word calendar description; (5) day-of reminder email. University voice: warm, no exclamation-point pileups, no invented quotes from the speaker.

Caution: Minimal, since this is public information by design. The failure mode is fabrication (an invented speaker credential, a wrong room), so the fact-check pass is the whole job.

Working looks like: Five channel-ready pieces from one paste, and a colleague can't tell which channel's copy came first.

For: Faculty

You need: ChatGPT Edu and lecture notes or a chapter you authored (or OER). The pattern comes from Ethan and Lilach Mollick's prompt library (CC BY 4.0).

Steps

1. Browse the Wharton prompt library's Instructor Aids and pick a pattern
2. Create a Project with your source material uploaded
3. Run the prompt against your material rather than the model's general knowledge
4. Review every question: is the marked answer correct, and is it in your material?
5. Fix or discard, expecting to keep 60 to 80 percent
6. Export the keepers to your quiz tool manually

Starting prompt

You are creating retrieval-practice questions from the uploaded material only. Generate 10 multiple-choice questions on [topic]: 4 plausible options each, exactly one correct, distractors reflecting real student misconceptions. After each, state the correct answer, why it's correct, and why each distractor is wrong, citing where in the material the answer lives.

Caution: Questions can be subtly wrong in discipline-specific ways; expert review of every item before students see it is non-negotiable. Don't upload publisher test banks; that's a copyright and an integrity problem at once.

Working looks like: A 10-question practice set where you signed off on each item, in a quarter of the usual time.