› Hi, Bot Games

Kitchen-table card games. Print tonight.

Screen-free decks that teach how bots work — the Sense / Think / Act / Power loop, training data, prompting, and debugging. Ungated PDFs — no email, no login. Tear on the perforation and play.

• 2–5 · ~15 min

  Loop Slap

  Slap the subsystem. Launch a bot. Take the chaos hit.

  The founding Hi, Bot card game. Reader speaks a Bot Builder hint; everyone slaps SENSE, THINK, ACT, or POWER. Collect three cards — Sense, Think, Act — shout BOT!, recite one If/Then breath, then slap Chaos: WEIRD, EDGE, or TRUST. POWER is an optional bonus; there is no Mission card. Checkable like Uno — the fun verb is the lesson.

  ~10–14 + parents · 6 PDFs

  Print Loop Slap →

• 2–4 · ~20 min

  Bot-Builders

  Install the loop. Survive the glitch. Reveal the bot.

  Set-collection for robotic engineers. Install five cards: one Sense, Think, Act, and Power part plus a Mission (Loop Slap needs three and has no Mission). Glitches hit the other benches. When someone shouts BOT!, last turns are install or discard only — then everyone reveals. Highest score wins — checkable like Uno, the verb is the lesson.

  ~8–12 + parents · 6 PDFs

  Print Bot-Builders →

• 2–4 · ~15 min

  Data Audit

  Draft the data. Face the tests. Find the blind spot.

  A drafting game about training data. Pick photos for your pet-spotting bot while Test cards reveal what the real world will ask — then a surprise Test nobody planned for. A Test passes only if one of your photos shows both of its words. Balanced data scores a bonus. Checkable like Uno: every point is on the table.

  ~8–14 + parents · 3 PDFs

  Print Data Audit →

• 2–4 · ~20 min

  Prompt Quest

  Say exactly what you mean. The robot does exactly what you say.

  A prompting game with Word cards instead of a keyboard. Line up words so a fetch-robot brings back the one Wanted thing — and nothing else. It reads literally, and a Quirk card bends how: short memory, a colorblind camera, mixed-up labels, trouble with NOT. Failed prompts stay on the table for the next player to refine.

  ~8–14 + parents · 4 PDFs

  Print Prompt Quest →

• 2–5 · ~20 min

  Debug the Bot

  Run the test. Rule it out. Name the broken part.

  A deduction game about troubleshooting. One player secretly holds the broken part; everyone else runs tests and reads the answers against a public Fault Chart. Name the fault in fewer tests for more points — so the best test is the one that splits the suspects in half. Sense, Think, Act, and Power, debugged.

  ~9–14 + parents · 5 PDFs

  Print Debug the Bot →

• 1–4 · ~25 min

  Matchbox Brain

  Teach a paper robot to win. Only tell it when it loses.

  A paper robot learns Hexapawn the way game-playing AI learns: by losing. It picks moves with a die from a printed Brain Mat, and every time it loses you cross off its last move in pencil. Nobody tells it why. It usually graduates within two dozen games, and once every losing move is crossed off it can never lose. The crossed-off mat shows exactly what it learned, and what it never saw.

  ~8–14 + parents · 4 PDFs

  Print Matchbox Brain →

• 2–6 · ~20 min

  Next Word

  Twelve cards, one die, a language model. Guess what it says next.

  A language model you can hold: twelve cards built by counting which word came next in a six-sentence story. Roll the die and it talks, one word at a time. Guess each next word, then vote on every sentence: remembered from the story, or made up? Most of what it says sounds right and never happened. That's a hallucination, and now you've made one.

  ~9–14 + parents · 3 PDFs

  Print Next Word →

• 1–4 · ~30 min

  Robot Route

  Program it for one maze. Teach it for every maze. Find where each one breaks.

  Act 1: write a program that walks a paper robot through a maze. It works, and it crashes on the very next maze. Act 2: teach the robot instead, by showing it what to do and tallying every move. Then it runs three mazes it has never seen, alone. One it solves, one it freezes on, and one it circles forever with the goal two squares away. Programs vs learning, and where each one breaks.

  ~9–14 + parents · 4 PDFs

  Print Robot Route →

• 1–2 · ~15 min

  Weight Pond

  Tune the frog's weights. Find out why one neuron isn't enough.

  A frog decides whether to jump with one sum: two weights and a bias. Drag the sliders until it lands on every lily pad with a fly, or press Let it learn and watch it fix its own mistakes. The last two ponds can't be solved by any single frog — you'll need two, and that's why neural networks have layers.

  ~10–14 + parents · 1 PDFs

  Play Weight Pond →

• 1–3 · ~20 min

  Guard the Bot

  Trick the bot into telling its secret. Then watch it get patched.

  Vault Bot is guarding a password, and every version has one weakness: a real trick people use on chatbots. Build messages from phrase cards, get the secret, then see the patch its makers add. Six vaults, from "just ask" to hidden instructions in a note. No real AI, no typing to a chatbot, and nothing you pick leaves the page.

  ~11–14 + parents · 1 PDFs

  Play Guard the Bot →

• 1–4 · ~25 min

  Twenty Questions Tree

  Ask yes or no. Sort the world. Then meet the objects your tree never saw.

  Build a tree of yes/no questions that sorts sixteen objects into Animal, Plant, Machine or Thing — the same way many real AI systems sort email and photos. Fewer questions score more. Then run six Surprise objects your tree has never seen through it. Coral, a robot dog, a wind-up car: a perfect tree on the training pile can still get them wrong.

  ~8–14 + parents · 4 PDFs

  Print Twenty Questions Tree →

• 1–5 · ~30 min

  Server Farm

  AI lives in a building. Keep it powered, keep it cool, keep up.

  AI isn't a cloud — it's racks of computers in a building. Run one for eight hours of growing demand: every rack needs power and makes heat, the sun sets at hour five, and the Event deck brings heatwaves and outages. Everyone shares the same dice, so the only difference is what you build. Racks, coolers, solar, batteries: keep up, or watch the jobs walk away.

  ~10–14 + parents · 4 PDFs

  Print Server Farm →

• 1–2 · ~25 min

  Bot Shell

  Type real commands into Pip's pretend computer. Find six hidden passwords.

  Every robot and every AI runs on an operating system. Pip the robot keeps hiding passwords on its computer, and you get them back by typing real commands: ls, cd, cat, grep, chmod, and pipes. Six levels, one new idea each. It's a pretend computer that lives in the page, so nothing can break and nothing you type leaves your device.

  ~12–16 + parents · 1 PDFs

  Play Bot Shell →

• 1–4 · ~30 min

  Rewind the Robot

  Train it, save every version, and roll back when the data goes bad.

  Train a helper robot one dataset at a time and save every version of its brain. Branch to try a risky dataset, merge two lines of work, and keep a log of exactly what went into each version. Then the Event cards reveal mislabeled photos, rude chat logs and data nobody agreed to share — and every version built with them changes. The winner ships their best version, which isn't always their newest.

  ~12–16 + parents · 4 PDFs

  Print Rewind the Robot →

• 1–4 · ~25 min

  Who's Important?

  Roll the dice, follow the arrows, and find out which spot in town really matters.

  A Surfer wanders a town by dice: follow one-way arrows, and now and then jump anywhere. Tally every spot it lands on, then guess which one wins. On Lakeside, the spot everyone points at loses to a spot with half the arrows, because importance comes from who points at you, not how many. It's PageRank, the idea that first ranked the web, played with two dice and a pencil.

  ~9–14 + parents · 3 PDFs

  Print Who's Important? →

• 3D · 1–2 · ~15 min

  Loss Mountain

  Walk the robot downhill in the fog. That's how AI learns.

  A 3D hill where height is how wrong a model is, and the robot can only feel the slope under its feet. Pick a step size and walk it down: too small and you run out of steps, too big and you fly off. 5 hills, from a simple bowl to a real one built from Weight Pond's pond 3. Walking downhill like this is how AI learns.

  ~10–14 + parents · 1 PDFs

  Play Loss Mountain →

• 3D · 1–2 · ~10 min

  GPU Paint Race

  One fast painter or 64 slow ones? Race them and see why AI runs on GPUs.

  Two 3D walls of pixels and two crews: one fast CPU painter, or 64 slower GPU painters who all do the same step at once. Pick who finishes first, then watch them race. 5 jobs, from painting a wall to following a trail where extra painters can't help, ending with a real neuron on a million dots. It's why AI runs on GPUs.

  ~10–14 + parents · 1 PDFs

  Play GPU Paint Race →

Want the screen-free activity library? Free printables · Start 14-day free trial

From sand to agents: 30 machines behind AI, to build for real →

Underfunded school, budget homeschooler, or family at home? See the whole free AI-literacy path.