COLL 100 · PROMPT ENGINEERING · HANDS-ON SESSION · IN CLASS, THEN ON YOUR OWN

“Hello, model.”

Your first lab. You'll run the same question through two different AI models, change one thing at a time, and watch what moves. By the end you'll have evidence — your own, logged — for Wednesday's sentence: fluent isn't the same as true.

GRADED ON DOING AND NOTICING Weird results are the good results.

You need: this page and your lab account — nothing else. The models run here; press RUN on any prompt box and the answer appears underneath it. Time: the hands-on session — done live, and the page saves as you go.

PRIVACYDon't put personal information — yours or anyone else's — into a prompt. Every prompt you run here is recorded with your lab.

SCOPEWe're not ranking models this week. We're learning to see differences.

1Your first run

GOALOne question, one answer from Model A, in your log.

REQUIREMENTSUse Prompt A exactly as written — this run is the baseline every later run is compared against.

DONE WHENRun 1 is in your log with the model named.

Every RUN on this page is independent: the model sees only the prompt in that box, with no memory of anything you ran before. That is what makes the comparisons in this lab fair.

Leave Model A selected in the box below.
Press RUN and read the answer that appears under the box:

PROMPT A

Explain in exactly 3 sentences why the same question can get
different answers from an AI. Write for a first-year student.

Press ADD TO MY LOG under the answer. It drops the prompt, the model's name and the exact words into your log in the submit panel at the bottom of this page.

✓

You should now see a 3-sentence answer in your log, with the model's name on it. Didn't happen? If the answer isn't 3 sentences — log it anyway. That's a finding, not a mistake: you just caught a model not following an instruction.

2Same prompt, second model

GOALThe same question, answered by a genuinely different system.

REQUIREMENTSSame prompt, character for character. Only the model picker changes.

DONE WHENRun 2 logged — two answers to one question, side by side.

Go back to the Prompt A box and switch the picker from Model A to Model B — a genuinely different model.
Change nothing in the prompt. Press RUN again.
Press ADD TO MY LOG. That is run 2, and the log records which model answered.

✓

You should now see two answers to the same question, from two different systems.

3Compare the two outputs

GOALA written comparison of the two answers.

REQUIREMENTSMark every disagreement — a fact, a number, the tone, the length — and say which one you would trust.

DONE WHEN1–2 sentences of your own in the Reflection box.

Both answers are now in your log, one under the other. Read them side by side and mark every disagreement — a fact, a number, the tone, the length. Type 1–2 sentences of your own into the Reflection box in the submit panel: what differs, and which one you would trust.

✓

You should now see two outputs that don't quite match. Different wording is normal. Different facts are the interesting part — flag those. Basically identical? Also a finding — log it; you'll probe it in Step 5.

4Rephrase it your way

GOALFind out what your wording changes.

REQUIREMENTSRewrite the question in your own phrasing. Keep the “exactly 3 sentences” rule or drop it — deliberately. Model A.

DONE WHENRun 3 logged, plus one line naming what your phrasing changed.

Edit the text inside the Prompt A box — same question, your phrasing. Keep the “exactly 3 sentences” requirement OR drop it deliberately.
Set the picker back to Model A and press RUN. Add it to your log — that is run 3.
Add one line to your log saying what your phrasing changed.

✓

You should now see how much (or little) the wording moved the answer.

5Run it twice more — watch the drift

GOALSee drift with your own eyes.

REQUIREMENTSThe original Prompt A text, Model A, twice more. Change nothing.

DONE WHENRuns 4 and 5 logged, plus one line: what drifted between three identical runs?

Put the original Prompt A text back in the box, Model A selected.
Press RUN, add it to your log, then press RUN once more and add that too — runs 4 and 5.
Same model, identical words in, three answers out (runs 1, 4 and 5). Add one line to your log: what drifted?

✓

You should now see that “I ran it yesterday” is not a guarantee about today. That's non-determinism — the same input can produce different outputs. Week 6 is all about it.

6Improve the weak prompt, three times

GOALTurn a genuinely weak prompt into a working one.

REQUIREMENTSThree passes. ONE labelled change per pass — task, context, or format — and a run after every pass.

DONE WHENThree labelled improvement passes in your log.

Here's a genuinely weak prompt:

THE WEAK PROMPT — EDIT IT, THEN RUN

tell me about william and mary

Make it better in three passes. One change per pass, and label the change with the part it adds — the working parts of a prompt: task (what to do, for whom) · context (what to work from, what to avoid) · format (the shape of the answer) · guardrails (what a good answer must include, or admit). Next week these become the five-part skeleton, with persona in front. Edit the box, press RUN, read the answer, then edit again for the next pass. Put all three passes — prompt, answer and the label of what you added — into the improvement passes box in the submit panel.

Example first pass (yours should differ): “Give a first-year student 5 facts about William & Mary that would matter in their first month on campus.” (adds: task + audience)

✓

You should now see the answer getting sharper each pass — and you can say WHY each time.

✎Journal · reflection · submit

Prompt journal: 3 real prompts you used this week for anything at all. One line each: did it work, and what one change would you make?
Reflection (5 sentences): Where did fluent and true come apart this week? Use at least one concrete example from your own logs.
AI Use Statement (50–150 words, required): which models you ran (A, B or both), that logged outputs are quoted verbatim (word-for-word, exactly as the model wrote them), what you fact-checked. Without it the submission comes back ungraded — that's the syllabus, not us being mean.

There is no file and nothing to upload: you do the lab right here in the hands-on session, and the submission panel below is where it all goes. Each piece is checked the moment you type it, and submitting records the lab and unlocks the next one.

HOW THIS IS GRADED — COMPLETION, IN ORDER
Completion, not polish — every item present and genuine, and the points are yours. Nobody grades your prose. The checklist: 5 logged runs (steps 1, 2, 4 and two in step 5) · 3 labeled passes · 3 journal entries · statement in range. Labs are 10% of the course, across six labs, and they unlock in order — one left undone blocks the next. No late window: finishing late beats not finishing — if you fall behind, tell your instructor rather than skipping ahead. AI use: Level 3 — submit your prompts and working results.

→Go further (optional, genuinely good)

OpenAI's Prompt Engineering guide — the pros' version of the five-part skeleton you meet next week.
Jakob Nielsen, “AI: First New UI Paradigm in 60 Years” — why typing intent is a big deal.
The 55-second film from Day 1 — watch it again now that you know what a prompt is. (Blackboard → Week 1 → “The film.”)

Why every RUN here is independent — the one-pager: A chat assistant rereads the whole conversation every time it answers. In a long chat, everything above influences what comes next — usually helpful, sometimes haunted: an old instruction (“answer in 3 sentences”) quietly shapes every later answer. Each RUN on this page starts clean instead. Rule of thumb: when you are comparing or testing, every run must start clean (which is why this page works that way); building on earlier work → stay in the chat, that's the point of it. If a model seems weirdly stubborn, check what's above — then start clean.