Prompt like
it actually matters.
Practice prompt engineering on real scenarios. Better prompts, fewer wasted tokens, less water and energy spent on slop. Built for anyone tired of vibes-based AI use.
has a cost
🔴 Lazy prompt
🟢 Considered prompt
How are these numbers calculated?
Tokens, multiplied by an estimated energy cost per token (around 0.0015 Wh from Patterson et al. 2021 and Luccioni et al. 2023), give the compute energy. Cooling water is a separate empirical estimate, around 500 ml per 250 tokens, based on AP News and UC Riverside reporting on data-centre water use.
Both numbers are indicative, not billing-grade. The point is the order of magnitude. CO₂ depends heavily on which grid is powering the data centre. Water is more direct, and lands harder.
Two tracks. One arcade.
Working pro or college student. Same engine, different scenarios. You can switch later, the receipts move with you.
Builder
Freshman
Bad prompts aren't just annoying. They're expensive.
Token spend
A vague prompt triggers 3 to 5 correction cycles. Each one re-sends your full context. Token cost compounds, even when each individual call is cheap.
Water per chat
A short ChatGPT conversation can cost ~500ml of cooling water. Microsoft's data-centre water use grew 34% in 2022. AI is a thirsty hobby.
Already happening
The fine one lawyer paid for citing AI-hallucinated cases in court. Recruiters, doctors, election callers, all sanctioned for AI output gone wrong. Real consequences, already here, not future tense.
When prompts go wrong, this is what it looks like.
Six real incidents. Each one starts with someone, somewhere, sending a prompt that lacked a constraint, a citation, a refusal, or a check.
The PromptArcade pledge
Every level you ship adds to your XP. But it also adds to a shared total: your tokens, your water, your CO₂. We surface this not to scold, but to make the cost visible. A well-prompted query is cheaper, faster, AND more likely to refuse the wrong thing. Craft is conservation.