ARQIV
Educational game · Prompt writing
Careful What You Wish For
Zafir the genie grants every wish exactly as written. Help eight customers word their wishes so nothing goes hilariously wrong, and learn the six ingredients of a clear AI prompt along the way.
Almost ready
The genie is still unpacking his lamp
The game is being installed and will appear here shortly. In the meantime, the guide below explains everything the game teaches.
Read the guide ↓Your progress saves on this device, even partway through a customer. Sound is on; switch it off in the game's Menu.
The guide · 8-minute read
The genie is literal. So is AI.
In the game, Zafir grants wishes word for word. Ask him for "a flyer for my bakery" and you get a perfectly nice flyer for the bakery. It just doesn't mention the gluten-free cupcakes, the school fair, or Saturday, because nobody said any of that.
AI assistants work the same way. They only know what you wrote, not what you meant. They can't see what you were picturing or what you'd call obvious. When something important is missing, they usually don't stop to ask. They fill the gap with a reasonable guess, and a reasonable guess tends to be a bland one.
In the game
Rosa the baker's wish, built from vague cards:
A flyer for my bakery, for families in town.
What happened: a general bakery flyer. Three parents ask where to buy the rye bread. Nobody asks about cupcakes.
In real life
The same request, typed into an AI assistant:
Write a flyer for my bakery.
What you get: "Fresh bread daily! Visit us today!" A cheerful flyer for a typical bakery, with nothing about the fair, the cupcakes, or who it's for.
Here's the same request with the missing pieces filled in:
Write a half-page flyer for my new gluten-free cupcakes. I'm handing it out to parents at Saturday's school fair, and many of them are looking for allergy-friendly treats. Mention the two flavors (chocolate swirl and vanilla berry), the price ($3 each), and where to find us (the booth next to the main entrance). Keep it under 60 words, upbeat but not cheesy. Don't say the cupcakes are safe for people with celiac disease, because my kitchen also handles wheat.
Nothing in the second prompt is clever. It just says what Rosa already knows. That's the whole skill.
Under the hood
How AI actually reads your prompt
The AI behind chat assistants is a large language model: a program trained on huge amounts of text to predict what comes next. Here's what happens to your prompt.
- It's split into tokens. A token is a small chunk of text: often a whole word, sometimes part of a word or a punctuation mark. In English, a token averages about four characters.
- The model predicts, one token at a time. Given everything so far, it picks a likely next token, adds it, and repeats until the answer is done. Modern assistants are also trained to follow instructions and be helpful, but every answer is still built this way.
- It only uses what's in its context window. The context window is all the text the model can use while it writes: your prompt, the conversation so far, any files you attached, and the answer itself. Think of it as working memory. Today's leading models can hold hundreds of thousands of tokens, and some up to about a million.
1 · Your prompt
"Write a flyer for my bakery."
2 · Tokens
An illustrative split. Real tokenizers vary.
3 · The model
Predicts a likely next token, adds it, and repeats, using only what's in its context window.
4 · The answer
A cheerful flyer for a typical bakery.
What the model can't see: the school fair is on Saturday · the cupcakes are gluten-free · the flyer is for parents · your kitchen also handles wheat. None of that was in the prompt, so none of it is in the flyer.
Anything outside the context window doesn't exist for the model. It has no access to your head, your situation, or the assignment sheet on your desk, unless you put it in. (Some apps add things for you, like saved memories or web search results, but only what actually gets added counts.)
So what happens to the gaps? The model doesn't leave them blank. It goes with whatever is most likely, based on the patterns it learned, and the most likely answer is the most typical one. Ask for "a flyer for my bakery" and you get the average bakery flyer. Vague prompt in, generic answer out. Laziness has nothing to do with it. Prediction is doing exactly what prediction does.
People often say a model "thinks" or "understands." It can be remarkably capable, but what it actually does is predict text based on patterns. Either way, it can't read your mind. Want the deeper version? See how transformers work.
A short history
How prompting has changed
The core skill hasn't changed: say what you want. But how people said it has changed a lot, because the models kept changing. Here's the short version, with one example request throughout: explaining how volcanoes form.
-
Roughly 2019 to 2021
Early completion models
Models like OpenAI's GPT-2 (February 2019) and GPT-3 (described in May 2020) were trained to do one thing: continue text. They didn't take instructions. So people wrote the beginning of the document they wanted and let the model finish it. A popular trick was few-shot prompting: show a couple of worked examples, then leave the last one blank so the pattern tells the model what to do.
Rewrite each science fact for a 10-year-old. Fact: Photosynthesis turns light energy into chemical energy. For a 10-year-old: Plants catch sunlight and turn it into food. Fact: Volcanoes form where melted rock rises through cracks in Earth's crust. For a 10-year-old:
Why it worked
The examples did the instructing. The model spotted the pattern and continued it.
-
2022
Instruction-tuned chat models
In January 2022, OpenAI described InstructGPT: models trained with human feedback to follow instructions. (The method is called RLHF, reinforcement learning from human feedback. People rank the model's answers, and it learns to give the kind they prefer.) When ChatGPT launched on November 30, 2022, anyone could simply ask. Two habits took off. Role prompts ("You are a…") set the voice. And the phrase "Let's think step by step," from a May 2022 study that built on Google researchers' chain-of-thought prompting, helped models work through harder problems.
You are an expert geology teacher. Explain to a 10-year-old how volcanoes form. Let's think step by step.
Why it worked
Direct requests finally worked. Roles and step-by-step nudges improved answers on harder questions.
-
2023 to 2024
The "prompt engineering" boom
2023 brought a wave of assistants: GPT-4 (March), Anthropic's Claude (introduced in March, with its public website opening in July), Google's Bard (March, renamed Gemini in February 2024), and Meta's open-weights Llama 2 (July). Developers got system prompts, standing instructions that sit above the conversation. Prompts grew long and structured, split into sections with delimiters (marks like triple quotes or XML-style tags such as <task>), and AI companies published official prompting guides. "Prompt engineer" became a job title, and lists of magic words that supposedly unlocked better answers spread online.
SYSTEM: You are a world-class science communicator for children. <task>Explain how volcanoes form.</task> <audience>10-year-olds</audience> <rules> - Exactly 3 short paragraphs - No jargon - End with a fun fact </rules>
Why it worked
Structure helped models keep long instructions straight. The magic words were hit-and-miss.
-
Late 2024 to today
Reasoning and agentic models
In September 2024, OpenAI released o1-preview, a reasoning model that works through a problem before it answers. Others followed quickly: Google's Gemini 2.0 Flash Thinking (December 2024), DeepSeek's open-weights R1 (January 2025), and Anthropic's Claude 3.7 Sonnet with extended thinking (February 2025). Models also learned to use tools like web search and code, and to keep going on longer, multi-step tasks. Open standards such as the Model Context Protocol (November 2024) made it easier to connect them to apps and data.
These models follow detailed instructions much better, so tricks matter less. OpenAI's own guide for its reasoning models says to keep prompts "simple and direct" and calls step-by-step instructions unnecessary. What matters more is giving the model the right material: a clear goal, the real context, your constraints, and examples. In mid-2025, people started calling this context engineering, a term popularized by Shopify CEO Tobi Lütke and AI researcher Andrej Karpathy. No magic words are needed. The job is filling the model's context with what it needs to succeed. That's exactly what WISHES trains.
I'm a 5th-grade teacher. Tomorrow I'm explaining how volcanoes form to a class that loves hands-on demos but struggles with long readings. Write a 3-minute explanation I can read aloud, plus one safe demo that uses kitchen supplies. Keep the explanation under 300 words. Don't use the term "tectonic plates" without explaining it first. My style is playful, like: "Imagine Earth is a giant peach with a thin skin." If anything is unclear, ask me before you start.
Why it works
No tricks, just a clear brief. What, Info, Shape, Hard limits and Examples are all there, and the last line invites the model to ask questions.
Side by side
Same request, three eras
Marcus (customer #2 in the game) needs to ask his very busy boss for Friday off, because his sister is getting married. Here's how people might have prompted for that message over the years.
Completion era
Subject: Friday Hi Dana, I wanted to ask whether
What you'd get: the model finishes the email, but it has to invent the reason, the tone and the length. It might ramble for paragraphs.
Chat era
You are a professional HR expert. Write an email to my boss asking for Friday off.
What you'd get: a stiff, formal email with blanks like [Reason] and [Your Name].
Prompt-engineering era
SYSTEM: You are an expert business communicator. <task>Request time off</task> <tone>Professional</tone> <length>Short</length>
What you'd get: a tidier message that's still generic. The model doesn't know why, which day, or that Dana reads everything on her phone.
Today, with WISHES
Write a short text message to my manager, Dana, asking for this Friday off. My sister is getting married that day. My quarterly report is already done, and Priya has agreed to cover my client calls. Dana is kind but very busy, and she reads messages on her phone between meetings. Keep it to three sentences or fewer. Don't over-apologize, and don't add details she doesn't need. Match this tone: "Hi Dana! Quick one when you have a sec."
- WWhat: a short text message to Dana asking for Friday off.
- IInfo: sister's wedding, report finished, Priya covering, Dana is busy and on her phone.
- SShape: three sentences or fewer.
- HHard limits: no over-apologizing, no extra details.
- EExamples: "Hi Dana! Quick one when you have a sec."
- SSharpen: read the draft, then fix what's off. "Good. Make the thank-you sound less formal."
Looking ahead
Where prompting is headed
Possibilities, not predictions
Nobody knows exactly how we'll talk to AI in five years. These are directions you can already see in today's tools. How far each one goes is still an open question.
Clarifying questions
AI that asks first
Like the question round in the game, some tools already ask before starting a big job. OpenAI's deep research, launched in February 2025, may ask clarifying questions before it begins. Expect more of this for long tasks.
Memory
Personal context
Assistants are starting to remember what you've told them across conversations; ChatGPT began testing memory in February 2024. That could mean less repeating yourself, and new questions about privacy and what you want an AI to keep.
Agents
Multi-step tasks
Agents plan, use tools, check their work, and keep going. When an AI acts for you instead of just answering, a clear brief (the goal, the limits, what "done" looks like) matters even more.
Multimodal
Beyond text
You can already prompt with photos, screenshots, voice and files. Pointing at a picture can replace a paragraph of description, but you still have to say what you want done with it.
Briefing
A colleague, not code
Prompting is starting to look less like writing code and more like briefing a capable colleague: explain the goal, share what you know, set boundaries, show an example, review the work.
However smart AI gets, one thing won't change: you are the one who knows what you want. The AI can't get that from your mind, only from your words. Saying it clearly will always matter.
Keep this handy
Your WISHES cheat sheet
Six ingredients, shown with one example task: getting help studying for a biology test. You won't need all six every time, but the ones you skip are the ones the AI will guess.
| W | WhatThe thing you want made. | Make me a one-page study guide. |
| I | InfoWho it's for, why, and what only you know. | It's for my 10th-grade test on cellular respiration this Friday. I get glycolysis but get lost in the Krebs cycle. |
| S | ShapeStructure and length. | Use 4 short sections with bullet points, then end with 5 practice questions. |
| H | Hard limitsWhat to avoid or leave out. | Stick to what's in chapter 7. Don't show answers to the practice questions until I ask. |
| E | ExamplesShow the style or tone you want. | Explain things like this: "The mitochondria is a power plant, and glucose is the fuel delivery." |
| S | SharpenRead the result, then fix what's off. | The Krebs section is too dense. Turn it into a 4-step list with one everyday analogy. |
For teachers
Using the game in class (about 40 minutes)
- Play in pairs (20 min). One student reads the customer's request aloud, the other builds the wish. Switch roles for each customer. Most pairs finish three or four customers.
- Discuss the funniest disaster (10 min). Which ingredient was missing? What would the same mistake look like with a real AI tool?
- Rewrite a real prompt (10 min). Each pair takes a vague prompt they've actually used, like "help me with my essay," and rewrites it with WISHES. If your school allows AI tools, try both versions and compare the answers.
How the stars work. Three stars for a wish with no flaws, on the first Grant or after one round of Sharpen. Two stars with one flaw left. One star with two flaws, as long as at least half the wish is right. Fixing a wish is part of the skill, so the game does not take stars away for it. A flawless first Grant with only two questions also earns a First try badge.
The cheat sheet above prints on one page. More lessons and printables: ARQIV for teachers →
Credits
Game by ARQIV. Art and audio created with AI. Built with open-source software: third-party notices and licenses.
Sources
- OpenAI, "What are tokens and how to count them?" help.openai.com
- Anthropic, "Context windows." platform.claude.com
- OpenAI, "Better language models and their implications" (GPT-2), February 2019. openai.com
- Brown et al., "Language Models are Few-Shot Learners" (GPT-3), May 2020. arxiv.org/abs/2005.14165
- OpenAI, "Aligning language models to follow instructions" (InstructGPT), January 2022. openai.com; paper: arxiv.org/abs/2203.02155
- Wei et al., "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," January 2022. arxiv.org/abs/2201.11903
- Kojima et al., "Large Language Models are Zero-Shot Reasoners," May 2022. arxiv.org/abs/2205.11916
- OpenAI, "Introducing ChatGPT," November 2022. openai.com
- OpenAI, "GPT-4," March 2023. openai.com
- Anthropic, "Introducing Claude," March 2023, and "Claude 2," July 2023. anthropic.com, anthropic.com
- Google, "Try Bard and share your feedback," March 2023, and "Bard becomes Gemini," February 2024. blog.google, blog.google
- Meta, "Meta and Microsoft introduce the next generation of Llama," July 2023. about.fb.com
- OpenAI prompt engineering guide (delimiters). platform.openai.com; Anthropic prompt engineering overview (XML tags). platform.claude.com; Google, "Prompt design strategies." ai.google.dev
- OpenAI, "Introducing OpenAI o1-preview," September 2024. openai.com
- TechCrunch, "Google releases its own 'reasoning' AI model," December 2024. techcrunch.com
- DeepSeek, "DeepSeek-R1 release," January 2025. api-docs.deepseek.com
- Anthropic, "Claude 3.7 Sonnet and Claude Code," February 2025. anthropic.com
- OpenAI, "Reasoning best practices." platform.openai.com
- Anthropic, "Introducing the Model Context Protocol," November 2024. anthropic.com
- Simon Willison, "Context engineering," June 2025 (quotes Tobi Lütke and Andrej Karpathy). simonwillison.net; Anthropic, "Effective context engineering for AI agents," September 2025. anthropic.com
- OpenAI, "Memory and new controls for ChatGPT," February 2024. openai.com
- OpenAI, "Introducing deep research," February 2025, and "Deep research FAQ." openai.com, help.openai.com