Camp curriculum · Project 5 of 5

Project 5: Tic-Tac-Toe

Kid line: You play tic-tac-toe against the robot. It looks at the board and places a mark.

Who does what

Younger (play)Older (copy / modify)
SkillPreloaded game. They play a real round. When it cheats or guesses, they name it — then keep playing.Copy the skill. Make it more robust: only empty cells, check before place, survive a trap board.

This is the finale. Skills 1–4 were one look → one job. Here the world changes every turn. Everything they already met shows up again: seeing, a recipe of places, guessing, confidence, and Stop.


Learning objectives (this skill owns)

  1. Reasoning over a changing world — Different board → different cell. Not one replay forever.
  2. Hallucination / illegal moves — It can “play” in a full cell or claim a win the picture does not show. That’s guessing.
  3. Not a person — It does not “want to win.” It matches a picture to a cell name.
  4. Robustness (older) — A wiser skill refuses bad moves: only empty cells, look again, Stay if unsure. Robust ≠ unbeatable. Robust = fewer illegal or invented moves.

What younger students should be able to say: “It guessed the next move from a picture of the board. It can guess wrong. We can still Stop.” What older students should be able to say: that, plus “We made it more robust. Empty cells only. Check before it places. It still isn’t perfect.”


Other learning opportunities (optional)

LensWhat you showKid takeaway
Confidence“You must win this turn” when that’s impossibleForced prompt → confidently illegal
Not a calculatorSame mid-game board, two runsDifferent cells is OK
Wrong cuePaper that says CENTER on a side cellText vs board (Skill 3 trick)
DemonstrationTap Cell 5 yourselfBlind body vs VLM eyes
Pick-up checkToken not in gripperSkill 2, on a game

Finale default: play a round + one illegal-move trap. Older: harden the skill, then the trap should fail (Stay or other empty cell).


The finished skill (instructors build this once)

Name in Raven: Tic-Tac-Toe Template: grid / pick-place per cell (Raven canvas grid if you use it).

Board

Motions (play skill)

ActionWhat to record
Cell 1Cell 9Pick a token → place on that cell → back to rest
StayNo place (unsure / illegal)

One token supply (small pile) in a fixed pick spot.

Prompt T — playable but not “robust yet”

This is what younger kids play. It will sometimes misread. That’s the finale, not a bug to hide.

You are placing X. Cells are 1–9, left to right, top to bottom. Look at the camera image. Choose one action: Cell 1Cell 9. Prefer an empty cell. If you are not sure, pick the lowest-number empty cell you can see. Do not claim you won unless three X’s are in a line in the image.

Runner (younger kids)

Younger kids do not edit prompts. They play.


Play (ages 7–10) — finale round

Keep it one short game, not a tournament.

  1. Empty board. Kid or instructor puts the first O (or robot goes first — pick one rule and stick to it).
  2. Each robot turn: Start. Arm places X. Kid places O.
  3. Narrate one turn: “It looked. It chose a cell. It replayed that place.”
  4. Trap (required): fill the cell it usually wants (often center). Robot’s turn. Legal = another empty cell. Illegal = goes to the full cell. Stop if it tries to stack. Class: “guessing” if it invented a free square.
  5. Finish the round or call time after the trap + two more turns.
  6. Ask them to explain it in their own words. Cheer the game, not “the AI is a genius.”

Got it: they can say that in their own words, and point to a legal vs illegal/guessed move. Almost: had fun; you named the miss. Not yet: “the robot is good at tic-tac-toe” with no look at the board.


Copy / modify (ages 10+) — make it more robust

Copy first: Tic-Tac-Toe (same nine places). Do not start from a blank grid unless the copy is broken.

Robust here means: fewer illegal places, clearer give-up. It does not mean a perfect player.

Required hardening (do at least two)

UpgradeWhat they changeWhy it’s robust
A. Empty-only promptReplace Prompt T with Prompt T-R (below)Model is told the rule, not just “prefer”
B. Pre-place checkVision validation before the chosen Cell runs, or a second look: Prompt T-VBody won’t place if the cell is full
C. Stay on unsureLow confidence / can’t see board → Stay, don’t guess centerGive-up is safer than a cheat
D. Trap testSame full-center (or full-corner) trap as playProof, not hope

They should fail the trap on the original copy first (or watch play fail it), then pass the trap after hardening.

Prompt T-R — robust order-taker

You are placing X. Cells 1–9, left to right, top to bottom. Look at the image. A cell is empty only if you can see it has no X and no O. Action must be one empty cell: Cell 1Cell 9. Never choose a full cell. Never invent a mark that is not in the picture. If no empty cell is clear, or more than one board is showing, action Stay. Do not say you won unless three X’s form a line in the image.

Prompt T-V — cell empty? (after they pick a cell, or as validation)

Is the target cell empty in the image (no X, no O)? If yes, pass and place. If there is already a mark, or you are not sure, fail — do not place. Action Stay.

Wire T-V as post-decision / pre-place validation if the Studio flow allows; otherwise they run a “Look” check then the Cell action only on pass.

What they must run

  1. Copy runs on an empty board (Cell place works).
  2. Trap on the unmodified prompt — log illegal or lucky-legal.
  3. Apply A + B (or A + C).
  4. Same trap again. Robust success: Stay or a different empty cell. No stack.
  5. Optional: “You must win now” on a board where that’s impossible — does T-R still refuse a full cell?
  6. Play one shortened game with the hardened skill so the finale still feels like a game.

Done when: trap is logged before and after, at least two upgrades are in the skill, and the older explanation.

Stretch: run the same board twice (W7); wrong-cue note CENTER; Skill 2 check on the token.

Not required: unbeatable strategy, both sides played by the arm, candy, barista.


Materials


If it fails

  1. Board unreadable → bigger marks, more light, camera higher.
  2. Always Cell 5 → favorite-cell greedy; hardening should add Stay / empty-only.
  3. Places off-grid → re-record that cell on this table.
  4. Illegal place in motion → Stop. Count it as the lesson, then harden.
  5. VLM down → kids tap cell buttons (blind). Finale still works as “we chose” vs “the model chose.”
  6. Don’t spend the finale in grid calibration. One working board per table is enough.

How this closes the camp

SkillWhat they can say at the end
1 CandyFace in the camera → a copied move
2 Intelligent DispenserCheck the gripper. Replay isn’t enough
3 RecycleLabels and confidence can fool it
4 BaristaA recipe is a list of copied pours
5 Tic-Tac-ToeA changing world, guesses, and a more robust recipe

Younger finale: play + trap + they can explain it. Older finale: same game, then they leave a sturdier skill behind.