Objective
Let students feel the gap between what they meant and what they actually said. One student becomes a "robot" who follows written instructions literally, with zero common sense. Teams learn that vague words like "draw a house" or "fold it" hide a hundred hidden assumptions — and that a good instruction leaves nothing to guess. This is the exact muscle they'll use to write sharp prompts for an AI agent.
The setup
- Split the room into teams of 3–4. Each team picks one robot — the person who will perform the instructions literally and silently.
- Pick one simple physical task for the whole class so teams can compare: draw a specific shape (e.g. a house with a door and two windows), fold a paper airplane, or stack 6 cups into a pyramid. Hand each team paper, a pen, and the materials.
- Post the Robot's Oath where everyone can see it: "I do only what the words say. If a step is unclear, I do the silliest thing that still technically obeys it. I never use common sense."
The rules — read aloud
Round 1 · "Say it loosely — watch it break"
- "Teams: write step-by-step instructions for the task on paper. You get 3 minutes. Write them the way you'd say them to a friend."
- "Robots: read your team's steps out loud and do exactly what each one says — no more, no less. If it says 'draw a house,' draw the wildest, most technically-correct house you can."
- "Teams may not talk, point, or rescue the robot. Zip it. Watch it go wrong and let it hurt."
Round 2 · "Rewrite it painfully specific — watch it work"
- "Same task. Now rewrite your instructions to be painfully specific and unambiguous — you get 4 minutes."
- "Add numbers, sizes, order, and positions: 'Draw a square 8 cm wide in the center of the page. On top of it draw a triangle of the same width for the roof…' Define every word a robot could twist."
- "Bonus: give the robot a quick example or reference ('like this picture') — good prompts often show, not just tell."
- "Robots perform again — literally. Whichever team's robot produces the correct result wins the round."
The process
- Round 1 (~5 min): 3 min to write, then robots perform. Expect chaos — houses with no walls, airplanes that are just crumpled paper, cup towers that fall. Let the room laugh; the failure is the lesson.
- Name it: "The robot isn't dumb — it's literal. Every gap you left, it filled with the silliest legal answer."
- Round 2 (~6 min): 4 min to rewrite tight, then robots perform again. Celebrate the team whose robot nails it. Ask them to read the one step that made the difference.
- Land it: "An AI agent is that robot. It has no common sense to fall back on — it does literally what your words say. Specific in, correct out. Vague in, garbage out."
The debrief
- Which vague word blew up Round 1? (e.g. "house," "fold," "middle," "some.") What hidden guess did the robot have to make?
- What did you add in Round 2 to make it work — numbers, order, positions, an example? Which mattered most?
- Why can't you just tell an AI agent to "make it look nice" and expect what's in your head?
- Finish this: "A good prompt leaves the agent ______." (nothing to guess · no room to misread · a clear format and example)
WORKSHOP TIE-IN: This is exactly why a sharp system prompt beats a vague one in today's Colab. Your agent is the human robot: be specific, define the output format, don't assume it "knows what you mean," and give it an example. You'll write two prompts for the same task — a loose one and a precise one — and watch the agent's answers split apart just like the two rounds.
WATCH-OUT: Keep robots playfully literal, not mean — the goal is funny-wrong, not humiliating. And cap the writing timers hard; teams will want to over-engineer Round 2. Specific-enough beats perfect, and the clock keeps the energy up.