“Write about data analysis.”
A generic, unfocused paragraph that could apply to anyone.
Output quality35%
No audience, goal, or format — the model guesses.
Tokenization
How models read text: tokens
Language models don't see words — they see tokens, the chunks text is broken into. Type below and watch the split happen; token count is what you actually pay for and what fills the context window.
Tokens0
Characters0
Chars / token—
Common words are one token; rare or long words split into several — roughly ¾ of a word each.
Next-token prediction
Predicting the next token
At its core a model just predicts the next token from a probability distribution. Temperature controls the gamble: low is focused and repetitive, high is creative and risky. Sample to see it choose.
Prompt: “The future of AI is ▮”
Higher temperature flattens the odds — unlikely tokens get a real chance, so output gets more varied.
Prompt engineering
Prompt Lab: good → better → expert
The same question gives wildly different answers depending on how you ask. Compare three levels of the same prompt and see what context, role, and constraints buy you.
“Explain data analysis to a marketing manager in 3 bullet points.”
Three focused bullets aimed at a marketer's concerns.
Output quality68%
Adds audience + format, so it's on-target.
“You are a data strategist. In 3 bullets, tell a B2B marketing manager how to use data analysis to cut CAC, with one metric each and a concrete first step.”
Role-grounded, metric-driven, actionable advice ready to use.
Output quality94%
Role + constraints + metrics + a first step = expert output.
Specificity, role, and constraints turn a vague reply into a usable one.