Why the Same Prompt Gives You a Different Answer Twice
It's not the tool being flaky. There's a dial doing this on purpose — and once you know it exists, you can control it.
You paste a prompt, get a great answer, paste it again, and get something different. People conclude the tool is unreliable. It isn't. It's sampling.
Temperature: the randomness dial
When a model predicts the next word, it doesn't just take the top candidate. It builds a probability distribution — this word 55%, that one 22%, another 9%. A setting called temperature controls how adventurous the pick is.
- Low temperature ≈ always take a top candidate. Consistent, repeatable, sometimes flat.
- High temperature ≈ sometimes take a less likely candidate. Varied, creative, occasionally unhinged.
Same model, same prompt, different dial position, different answer. That's the whole mystery.
When to turn the dial down
Anything you'll run repeatedly wants low variance: data extraction, classification, formatting, customer replies that must match a template, code that must follow a spec. You want the boring correct answer every time.
When to turn it up
Ideation, naming, angles, subject lines, first drafts where you plan to pick the best of ten. High temperature is a brainstorming partner that isn't attached to its own ideas.
The underrated third option
Rather than chasing perfect prompts, ask for three versions at low temperature, or a single version plus its two strongest alternatives. You get range without losing reliability — and you stop re-rolling the dice hoping for a better answer.
Control over this dial is one of the first things we drill in Prompt Foundations. It's the difference between a slot machine and a tool.
From trivia to income
Secure your $5/month access before the window closes.
AI Courses Bundle turns curiosity into capability: weekly course drops, practical systems and the AI Consultant Course business in a box.