• How to explain what AI actually is, in one plain sentence.
• The difference between AI, machine learning, and generative AI — without the jargon.
• Where AI is a strong fit for a task today, and where it still needs a human.

Generative AI tools like ChatGPT don't think or reason the way a person does — under the hood, they're next-word predictors: given the text so far, the model calculates which word is statistically most likely to come next, based on patterns it learned from huge amounts of text. That single mechanic explains a lot of what AI is good at (drafting, summarizing, brainstorming) and where it struggles (exact numbers, very recent events, anything that needs real judgment) — and it's why sounding confident is not the same as being correct.

Because generative AI works by predicting likely patterns from what it's seen before, it tends to be strong wherever there's a lot of precedent to draw on — drafting, summarizing, brainstorming, rewriting tone. It's weaker wherever a task needs a guaranteed exact fact, or a judgment call with real consequences for a person — those need a human in the loop, either to supply the missing accuracy or to own the decision.

Terms like "large language model," "hallucination," or "training data" get thrown around a lot, but they describe fairly simple ideas once you strip the jargon out. Practicing a plain-English version of these terms — the kind you'd use with a coworker who's never touched AI — is what actually makes the concepts stick, and what lets you explain AI's real strengths and limits to your team without sounding like you're reading a manual.

• How to describe, with a simple analogy, how an AI model actually generates a response.
• Why AI sometimes gets things confidently wrong — in one plain sentence.
• What a "context window" and "training data" mean, well enough to explain them to a colleague.

A large language model doesn't compose a whole sentence at once — it generates one token (roughly a word or word-piece) at a time, each time asking "given everything so far, what's the most likely next token?" It actually computes a probability for every possible next word and picks from the top of that list. This game shows you that real list — the model's actual top 5 guesses and how confident it was in each — so you can see next-word prediction happening, not just hear it described.

AI hallucination is when a model states something false with the same fluent, confident tone as something true — because it's not lying on purpose, and it's not checking facts against a database. It's generating the next most plausible-sounding words, and a fabricated detail can be just as statistically "likely" as a real one. That's why you can't catch a hallucination by ear — you have to actually verify the claim.

An AI model doesn't have a memory in the way a person does — everything it "knows" about your conversation has to fit inside its context window, a limited amount of text it can look at for that response. Once something scrolls out of that window, it's genuinely gone unless someone restates it — the model isn't ignoring it or forgetting on purpose, it just can't see text that's no longer in its window. That's why long AI conversations can quietly lose track of details from earlier on.

• The four main categories of AI tool — chat assistants, embedded AI, specialist tools, and search-enabled AI — with an example of each.
• How to pick the right category of tool for a given work task.
• How to set up an AI account with sensible memory and privacy settings.

• Chat assistant — a general thinking and writing partner (Claude, ChatGPT, Gemini). You paste in what it needs.
• Embedded AI — AI built into a tool you're already using (Copilot in Word, Excel and Outlook; Gmail's AI features). It works where you are.
• Specialist tool — built for one job, like meeting notes, slide decks, or images and voice.
• Search-enabled AI — an assistant that browses the live web and cites what it finds, so it can handle anything recent.

An AI tool without web access only knows what was in its training data, which stops at a knowledge cutoff. Ask it about yesterday's news or a live price and it can still answer fluently, but the answer is a guess. A tool with web search can look things up and cite where it found them, which fixes the staleness. It doesn't make the answer automatically right, though: it can pick a weak source or misread a good one, so anything important is worth a click to the original. Your three options are Trust it (fine to use as is), Verify first (plausible, but check a source before relying on it), and Throw it out (the tool couldn't have known, so ask one that can). Verifying when the answer was unusable earns half credit: safe, but you'd still have wasted the check.

Real AI tools ship with defaults that suit the vendor, not your situation. Memory carries details from one chat into the next, which is handy for a brand voice and risky for salary data. "Improve the model" settings let the provider learn from your conversations. Shared links can end up findable by anyone. Connectors give the AI a window into your email, calendar and files. The right setting depends on what you're about to put into the tool, so there is no single "safe" configuration. Some switches are marked your call in the results because either choice is defensible.

Every AI vendor wants their product to sound like it replaces all the others. Your job is to look past the pitch. Good questions are concrete: Where does it get its information? Does it work inside the app I already use? Can it look up today's news? What does it need from me to do a good job? The salesman is a real AI playing the part, so his answers are spin-heavy but he can't flatly lie about what the product is. Your score comes from the verdict: 60 points for naming the right category, 40 for spotting the real limitation.

• The 5-part prompt structure: Role, Context, Task, Format, Constraints.
• How to rewrite a vague prompt into a specific one, using the real facts of your work.
• How to tell exactly which part of a prompt is missing when the AI's answer disappoints.

• How to refine an AI output over several turns instead of starting over each time.
• Why asking an AI to show its work lets you catch its mistakes.
• How personas and examples (few-shot) lock in a tone or house style.

• Role — who the AI should be (“a senior sales director who writes warm, concise emails”).
• Context — the real background: names, dates, numbers, what happened. Paste the facts in.
• Task — the one specific thing you want done.
• Format — the shape and length of the answer (“subject line plus three short paragraphs, under 130 words”).
• Constraints — tone, things to avoid, and a guardrail like “only use the facts given; say so if you're unsure.”

When a reply disappoints, the cause is almost always one missing ingredient, and the symptom points to it. Generic voice? No role. Bland, could-be-anyone content? No context. The AI asks what you want? No task. A wall of text? No format. Invented facts or promises you can't keep? No constraints. Length limits count as format; tone, things to avoid and guardrails count as constraints.

The AI can only use what you give it. A vague prompt gets a generic answer that misses every specific requirement. To win the duel, put the real facts into your prompt (paste them in rather than describing them), say exactly what shape and length you need, and state the requirements as instructions. Word limits allow 10% slack, because even good prompts don't produce an exact count.

Treat an AI like a conversation, not a vending machine. Instead of throwing the draft away and re-prompting from scratch, give targeted follow-up instructions (“cut the second paragraph”, “lead with the number”, “no contractions”). The AI keeps everything you didn't ask it to change. Bundling several changes into one message saves turns, but watch for changes that undo each other.

Few-shot prompting means showing the AI one or two examples of what you want instead of only describing it. It copies whatever the examples do, good or bad, so a rambling example teaches rambling. A persona (“act as a concise operations lead reporting to a COO”) shifts the voice and priorities of the answer, and the wrong persona pulls it off-style. The five style checks after each run show exactly what your choices produced.

Asking an AI to “think step by step” before answering does two things: it tends to improve accuracy on multi-step problems, and it makes the reasoning auditable. A bare number gives you nothing to check, but numbered steps let you find exactly where a slip happened (a wrong percentage base, a unit conversion, a misread sentence). Fluent reasoning isn't the same as correct reasoning, so check the steps that matter. Some solutions are sound, so don't cry wolf on all of them.

• How one clear tone instruction turns a rambling, over-apologetic draft into a confident message.
• How to edit an AI draft instead of sending it as-is: spotting filler, invented facts and broken promises.
• How to rewrite one source for several audiences: client, executive and slide bullets.

The base request is fixed: it carries the facts but says nothing about tone. You write only the tone and style instruction that gets added to it. Good tone instructions name the register (“confident, brief, solution-focused”), set a length, and say what to avoid (“no over-apologizing, no clichés”). Both versions run live, and every requirement is checked automatically. Word limits allow 10% slack.

AI drafts are fast but not finished. Compare each sentence with the notes. Three kinds of problem turn up again and again: invented facts (numbers or promises that aren't in the notes), filler (empty clichés and puffery), and wrong tone or broken promises (over-apologizing, threats, commitments the company can't keep). Everything starts as “Fine”; flip a sentence to “Flawed” if it needs editing. Wrongly flagging a good sentence costs you a point.

The same facts need different clothes: a client wants reassurance and no internal jargon, an executive wants the outcome and the decision in two sentences, and a slide wants three short bullets. Tell the AI who the reader is, how long it should be, and what to leave out. The goals for each version are listed under its box and checked automatically. Word limits allow 10% slack.

• How to check an AI summary against the source and catch distortions and invented claims.
• How to clean a messy spreadsheet in plain English, and how specific the instruction has to be.
• Why you verify every citation before using a claim externally.

A summary can fail in two ways. It can distort the source (a number changed, a limit dropped, a detail overstated), or it can add something that is not in the source at all. Both sound just as confident as a correct claim. Compare every claim with the source text, word for word, before you trust it.

The AI can clean data well, but only as well as you describe the target: which date format, how names should look, what to do with duplicates and blanks, and what a number column should contain. Ask for CSV only so the result is easy to paste into a spreadsheet. The goal list shows what “clean” means here.

AI can write citations that look perfect and are wrong in three ways: the source says something different from the claim, the claim is right, or the source doesn't exist at all. The only fix is to look the source up. For each claim you see what a check of the source found.

• How to turn messy notes into action items with owners and deadlines, without letting the AI guess who volunteered.
• How to use AI for a pre-mortem: challenge questions that stress-test a plan without making the decision for you.
• Which jobs to hand to AI, which to draft-then-confirm, and which stay human.

Meeting notes are messy, and AI will happily tidy them into something that looks authoritative, even when the notes never said who owns an item. Tell the AI the format you want (owner, task, deadline), and tell it what to do when an owner or date is unclear: flag it, don't guess. Word limits allow 10% slack.

A pre-mortem asks “it's six months later and this failed: why?” Good prompts give the AI the whole plan, ask for challenge questions (not answers) across distinct kinds of risk, limit the list so you can actually use it, and forbid a verdict. You make the call. The AI is the devil's advocate. Word limits allow 10% slack.

Three buckets. AI can just do it: low stakes, and you'll see the result anyway. AI drafts, I confirm: AI does the heavy lifting, but you check facts, owners and numbers before it leaves your hands. Only a human: a judgment call, a people decision, or something the AI cannot know (who really volunteered, how the room felt).

• How to describe a picture precisely: subject, style, palette, composition and what to leave out.
• How to write a voiceover script that sounds right when read aloud by a text-to-speech tool.
• Which tool category to use (image, video or voice) and when a human must step in.

Image tools fill every gap with their own defaults. A strong image prompt names the subject (who or what, how many), the style (flat illustration, photo, 3D), the palette, the mood and light, the composition (angle, space left for text), and what to leave out. Both prompts are given to the same AI, and the description is checked against the brief. Word limits allow 10% slack.

Image tools make graphics, illustrations and photo-style pictures. Video tools make short clips and can enhance or upscale existing footage. Voice tools turn a script into narration. A human is needed for exact brand assets, factual accuracy of on-screen figures, real people's likeness, and usage rights. These are limits no tool removes, and Module 14 covers them in depth.

Text that reads well on a page often sounds terrible aloud. A voiceover script needs short sentences, no bullets or headings, no symbols or abbreviations a voice would stumble over (&, @, %, e.g.), spelled-out days and numbers where it matters, and a length that fits the time slot (about 130 to 150 words per minute). Word limits allow 10% slack.

• How to chain prompts, where each step's output becomes the next step's input.
• How to pick between one prompt, a chain, a saved assistant and an agent.
• How to set up a custom assistant with your voice, template and a rule for missing information.

A chain turns one big job into focused steps: for example calculate, then interpret, then write. Each step sees only what the previous step produced, so it must pass along what the next one needs. Your final step is also run alone on the raw data, so you can see what the earlier steps actually added. For small tasks, one prompt is often enough, and the game tells you when that is true.

One careful prompt: a single, self-contained job. A prompt chain: several distinct steps where you want to check each result. A custom assistant: the same setup (voice, template, reference material) used again and again, possibly by other people. An agent: an open-ended goal where the AI uses tools and takes several actions with less steering, and a human still checks in before anything irreversible.

A custom assistant (a “custom GPT”) is just a chat with instructions pre-loaded: who it is, the template it must follow, the voice to use, and what to do when something is missing. Paste the brand voice and template into your instructions, and tell it to say so rather than invent when data is absent. The left side runs the same request with no setup. Word limits allow 10% slack.

• The three ways generative AI creates real value: scale, less busywork and personalization.
• Why pilots fail: no clear problem, garbage in, or autopilot with no human review.
• How to make a business case a skeptical CFO would take seriously.

Generative AI creates value in three repeatable ways. Scale: more content or output without a proportional rise in headcount. Less busywork: repetitive tasks removed so people spend time on judgment-heavy work. Personalization: tailoring communication to each person at a level that used to be impossible. And some “projects” create no clear value: a tool with no problem to solve.

The same technology can win or fail depending on the setup. Three failure patterns recur: no clear problem (a solution looking for a problem), garbage in, garbage out (poor, outdated or biased data), and autopilot (nobody reviews what the AI does before it reaches customers). A healthy pilot has a defined goal, a baseline to measure against, and a human reviewing the output.

A good business case answers four questions. What problem and goal? (specific, not “explore AI”) What is the measurable value? (a baseline and an expected gain, with numbers) Where is the human? (who reviews the output before it goes out) Is it worth it? (an honest view of cost, effort and disruption, including what could go wrong). The CFO is an AI scoring each answer from 0 to 25 against fixed anchors, so specific beats impressive.

• How to write a solution-neutral problem statement that states the gap and its cost, without smuggling in a tool.
• The four-step opportunity framework: Discover, Define, Develop, Deliver.
• How to size an opportunity honestly: how often it recurs and how painful it is.

A good problem statement describes the gap and what it costs, in plain terms, and leaves the answer open. “We need a faster way to review supplier contracts for standard clauses” works. “We need a chatbot that reads PDFs” doesn't, because it smuggles in a tool. Include the numbers (time, volume), who is affected and where the friction is, in one or two short sentences.

Discover: map how the work really happens today, who is involved and where the friction is. Define: turn what you found into one clear, solution-neutral problem statement. Develop: generate two or three different possible approaches, not just the first idea. Deliver: pick the strongest option and define what success looks like and what you need to try it safely.

Not every friction point deserves a fix. Multiply it out: time × people × how often, and weigh how painful and error-prone the task is. As a rough guide, hundreds of hours a year with real pain is a strong candidate; around 50 to 200 hours, or lots of hours with little pain, is marginal; a handful of hours or a one-off is not worth the effort, since setting up and checking an AI solution has its own cost.

• How to spot the warning signs of a hallucination: invented specifics, fake sources and overconfidence.
• The difference between verification (did I get what I asked for?) and validation (does it solve the real problem?).
• How to prompt for honesty so the AI admits what it doesn't know.

AI is confidently wrong: it doesn't know when it's guessing. Watch for invented specifics (a detail, quote or fact that wasn't in what you gave it), fake or unverifiable sources (precise citations you can't find), and overconfident certainty (“always”, “proven”, “no risk”). A claim that can be checked directly against the material you pasted is far safer, but you still check it.

Verification asks “did I get the thing I asked for?”: the right format, length and facts. Validation asks “is it the right thing?”: does it solve the real problem, not just satisfy the prompt. An output can pass verification and still fail validation, such as a summary that is well formatted but misrepresents the real person. Both checks matter, and they catch different failures.

A model that isn't told otherwise will usually produce a confident, plausible answer. You can change that. Tell it to say when it doesn't recognize something or can't know it, to never invent figures, names or sources, and to point you to where you can check. Both versions run live. Word limits allow 10% slack.

• What should never be pasted into a public AI tool, and what can be made safe first.
• How AI has changed scams: voice cloning, deepfakes and perfectly written phishing.
• The habit that defeats most of them: verify unusual requests on a second, independent channel.

Public AI tools may store or learn from what you paste. Never paste client or company confidential information, credentials, financial results before release, or sensitive personal data. Anonymize first when the task is fine but the text contains names, contact details or figures you can remove. Fine to paste covers public material and your own generic work. Approved enterprise tools may have different rules, so follow your organization's policy.

AI now clones voices, imitates writing styles and fabricates faces on video, so how convincing it sounds is no longer evidence. Requests that involve money, credentials or changed instructions, even from someone you know, need verification on a second, independent channel (call back on a known number). Unexpected links, attachments or login pages from unfamiliar senders should be reported to security without engaging. Genuinely routine requests proceed normally.

Take out everything that identifies a person, company, account or amount, and keep what the AI needs to do the job: the task, the situation in general terms, and the tone. Replace specifics with neutral descriptions (“a client”, “an invoice over 30 days late”). Don't just delete everything, because a useless prompt is also a failure.

• How humans and AI divide the work: people own judgment and accountability, AI handles volume and first drafts.
• Which skills outlast any tool, and which of yours to protect from over-reliance.
• How to talk about AI anxiety honestly, and how to stay current with a small, steady habit.

The healthy pattern is humans and AI dividing the labor. Humans own judgment, accountability and ambiguous or sensitive situations. AI drafts and a human decides covers volume, first passes and anything that goes out under your name. AI can handle low-stakes, repetitive or throwaway work, where you only skim the result. Think “co-pilot, not autopilot”.

Tools get replaced every year or two, but habits transfer: specific prompting, verifying, choosing the right tool. Knowledge of one tool's menus and buttons expires when the tool does. And some skills are at risk from over-reliance (writing fluency, estimating, critical reading, negotiation instinct) and need a deliberate safeguard so they don't fade.

Each situation has its own checklist, shown before you write. Good answers are specific (name the skill, the cadence, the step), honest (neither “nothing will change” nor “we're doomed”) and about what is in your control.

• How to write a reusable prompt: a placeholder for the input instead of one-off details.
• The documentation standard: task, prompt, expected output, how to verify, how to validate, and when not to use it.
• How to turn a rough idea into a one-page pitch someone could actually approve.

Keep the six labelled lines: Task:, Prompt:, Expected output:, How to verify:, How to validate: and When not to use:. Put each label at the start of a line. A reusable prompt has a [placeholder] where the input goes, and it never contains real names, emails or ID numbers.

Each entry has six parts: the task, a reusable prompt, the expected output, how to verify it, how to validate it and when not to use it. Plain rules score every entry out of 100, and your game score is the average of your best five entries, so five strong entries earn full marks. Your entries are saved to your profile and can be downloaded at any time. You can score the game up to 3 times.

Use five labelled lines: Problem:, Cost today:, Proposed approach:, Success looks like: and What I need to try it safely:. State the problem without naming a tool, put numbers on the cost and on success, keep a human reviewing the AI's work, and don't over-promise.

• How to give a two-minute lightning pitch that a busy audience can follow.
• How to build a 30-day follow-through plan on top of the staying-current habit from Module 15.
• A final assessment across the whole course. Pass it to earn a printable DaRI certificate.

• Replay any game to beat your best score. You get three scored attempts on each.
• Keep building your playbook from Module 16, and download it whenever you add to it.
• Follow your 30-day plan. A short, regular check-in beats a big push once in a while.
• Ask your facilitator about advanced topics when you are ready to go further, and about where to get support after the program.

A two-minute talk is about 200 to 300 words. Open by saying what it does, share real figures, give one example, be honest about the human check or the limit, and end with one clear ask. Write it for the ear: short sentences, no bullets, no symbols.

A plan that sticks is small and specific: one action for each of the four weeks, a short regular check-in with a time limit, one skill you'll protect from over-reliance, one person or place to ask for help, and one number you'll aim for.

The questions are drawn from every module and change each time you play. Some come with a short source or situation to read first. There's no time limit. You need to have played at least one game in each of Modules 1 to 16 before you can sit the exam. You get three scored attempts, and your best score counts.

Study the picture above closely, then describe it in the prompt box exactly as you see it — objects, colors, style, composition. Submitting generates a new picture from your words, and an AI judge scores how closely it matches the original, out of 100.

An AI art judge compares your recreated picture to the hidden original across several dimensions: the exact subject (and how many of them), the color palette, the composition and framing, the art style and medium (photo vs. painting vs. cartoon, etc.), the setting and background, and small distinguishing details like lighting, texture, or accessories. It's deliberately strict and uses the full 0–100 scale, so scores in the 90s are reserved for near-perfect matches — even one clearly wrong detail (a different color, a missing prop) costs several points. It never judges how good your picture looks on its own, only how closely it matches the original.

Both players see the same hidden picture each sub-round, on their own device. Type your prompt whenever you're ready — you don't have to wait for your opponent to go first. Once you've both submitted, each device generates and scores its own player's picture, so neither of you sits around waiting on the other's turn. Whoever scores higher that sub-round wins it — the match goes to whoever's total score is higher across every level picked, with a tiebreaker sub-round if it ends up tied.

An AI art judge compares your recreated picture to the hidden original across several dimensions: the exact subject (and how many of them), the color palette, the composition and framing, the art style and medium (photo vs. painting vs. cartoon, etc.), the setting and background, and small distinguishing details like lighting, texture, or accessories. It's deliberately strict and uses the full 0–100 scale, so scores in the 90s are reserved for near-perfect matches — even one clearly wrong detail (a different color, a missing prop) costs several points. It never judges how good your picture looks on its own, only how closely it matches the original.

Read the emoji string — the category badge next to it tells you what kind of answer you're looking for (for example a movie, a song or an idiom). Type your best guess below — you can reveal a clue if you get stuck. Scoring is instant and compares your guess directly to the hidden answer, out of 100.

Scoring is done locally, without any AI call. Your guess is compared to the hidden answer two ways: how many of its key words you got right (ignoring small words like "a" or "the"), and how close your spelling and phrasing are letter-for-letter. The two are blended — weighted a bit more toward key words — into a score out of 100. An exact match is always 100; a guess that shares no words or letters with the answer scores near 0.

You and your opponent each play on your own device. Every sub-round you both see the same emoji string and guess independently — neither of you can see the other's guess until you've both answered. Whoever's guess scores higher wins that sub-round, and the match goes to whoever's total score is higher across every level, with a tiebreaker sub-round if it ends up tied.

Your guess is scored the moment you submit it, without any AI call. It's compared to the hidden answer two ways: how many of its key words you got right (ignoring small words like "a" or "the"), and how close your spelling and phrasing are letter-for-letter. The two are blended — weighted a bit more toward key words — into a score out of 100. An exact match is always 100; a guess that shares no words or letters with the answer scores near 0.

Read the hidden Python function above and figure out what request would have produced it. Describe that request in the prompt box — it generates a new function from your words, and an AI reviewer scores how closely its behavior matches the original, out of 100. Nothing is ever run.

An AI code reviewer reads your recreated function side-by-side with the hidden original — nothing is ever executed. It judges whether the two would behave the same way on typical and edge-case inputs (empty input, negatives, duplicates, and so on), whether they take a similar algorithmic approach, whether the function's signature and overall structure match, how robust each is (error handling, input validation), and how similar the readability and style are. It's strict: scores above 90 are reserved for functionally equivalent code, while a behavioral mismatch or a completely different approach to the problem scores below 20.

You and your opponent each play on your own device. Every sub-round you both see the same hidden function and write your own prompt to recreate it — neither of you can see the other's prompt or function until you've both been scored. Whoever's function scores higher wins that sub-round, and the match goes to whoever's total score is higher across every level, with a tiebreaker sub-round if it ends up tied.

An AI code reviewer reads your recreated function side-by-side with the hidden original — nothing is ever executed. It judges whether the two would behave the same way on typical and edge-case inputs (empty input, negatives, duplicates, and so on), whether they take a similar algorithmic approach, whether the function's signature and overall structure match, how robust each is (error handling, input validation), and how similar the readability and style are. It's strict: scores above 90 are reserved for functionally equivalent code, while a behavioral mismatch or a completely different approach to the problem scores below 20.

Edit the broken prompt below so it satisfies the requirement listed above, then click "Test repair" to run it against the one real message shown. You get 3 attempts — editing is free, only testing uses one up, and a technical error never costs an attempt. Hints cost 10 points each. Solving it sooner scores higher.

Your edited prompt is run once against the customer message shown, and the AI's response is checked against the requirement. Solving it on attempt 1 scores 100, attempt 2 scores 75, attempt 3 scores 50 — not solving it in 3 attempts scores 0. Each hint you use subtracts 10 points from whatever you scored.

Both players repair the same ticket with their own private 3 attempts and hints. Once both are done, whoever scored higher wins that sub-round — the match goes to whoever's total score is higher across every level you picked, with a tiebreaker sub-round if it ends up tied.

Each player's edited prompt is run once against the customer message shown, and the AI's response is checked against the requirement. Solving it on attempt 1 scores 100, attempt 2 scores 75, attempt 3 scores 50 — not solving it in 3 attempts scores 0. Each hint used subtracts 10 points from whatever that player scored.

• Work through 5 sequential stages of one continuous mystery — read the evidence for each stage, then ask Vex, your AI detective partner, or interrogate suspects directly to dig deeper.
• Submit your conclusion for each stage. A wrong answer costs points but never locks you out — you can always try again.
• Hints are available any time, in 3 tiers, but each tier costs more points the further it goes.
• Your progress is saved automatically on this device, so you can leave and come back to the same case.

Vex checks your stage conclusions and final solution against what the evidence actually supports — it accepts differently worded answers that convey the right idea, not just exact keywords. Your final score starts at 1000, adds a time bonus for finishing inside the time limit, and loses points for hints used and wrong submissions.

• Pick a mission. You'll always see which objects are available and what each goal looks like — nothing is hidden.
• Type or speak your instructions to Robo, then watch what it does.
• You get 3 tries per mission — improve your instructions each try. Once a goal is done, it stays done.
• Earn a star for every goal you complete clearly. No points are ever lost for spelling, grammar, accent, or short sentences.
• There is never a timer — take as long as you like. In a tournament, you and your opponent each play the same mission on your own devices: the most stars wins, and if the stars are level, whoever needed fewer tries.

Robo only ever exploits a real gap in your instructions — never a good, clear one. If your instructions are already clear, Robo just does the mission right the first time. It's never trying to trick you or make you feel silly.

You and your opponent each play on your own device. Every mission, you both get the same scene and the same objects. Tell Robo what to do, up to 3 tries, and earn a star for every goal you get done clearly. Neither of you can see the other's instructions or stars until you've both finished. The match goes to whoever has more stars in total — if the stars are level, whoever needed fewer tries — with a tiebreaker mission if it is still level. There is no timer: take as long as you like.