Why answers vary

Send the same prompt twice, get two different answers. Nothing broke: a model chooses each next token by sampling from a range of plausible options — not by looking up one fixed reply. Small differences — your wording, the order of the conversation, the product’s settings — shift which option gets picked.

Where the variance comes from

• Sampling settings. Products control randomness with parameters like temperature: lower is more predictable, higher is more varied. Many apps set these for you and never show them.
• Tiny prompt changes. One added word nudges every probability that follows. “Make it shorter” and “keep it brief” are not identical inputs.
• Context changes. The same question three turns later sits next to a different history.
• Model versions. Updates change behaviour without changing anything you did.

What varies — and what that is good for

Variance is a feature for brainstorming: ten runs, ten angles to choose from. It is a hazard for anything that must stay stable: legal text, official numbers, approved copy. The working rule: drafting tolerates variance; commitments do not.

A bad example

Expect: “Ask once, approve it, ask again tomorrow and get word-for-word the same text.”

You won’t. Something shifts — wording, structure, an example. If you promised the approved version, you no longer control what was promised.

A better example

Workflow: generate until you approve a version, then save that text and work from the saved copy. When variance is acceptable, re-run for alternatives; when it is not, paste the locked text back in.

Why it works

Treat model output as a draft with variance, not a deterministic function:

1. Snapshot approvals. The file you saved is the commitment; the chat is not.
2. Re-run for diversity, not confirmation. Need alternatives? Ask for them on purpose.
3. Don’t chase exact repeats. “Say it exactly like before” has nothing to retrieve — unless the previous text is in front of it.

The practical consequence: keep prompts like recipes, not incantations. Save the version that produced a good result, note what you changed when it stops working, and treat every run as a sample rather than the model’s single true answer. Two samples that agree are stronger evidence than one sample you liked — that is the whole habit in one line.

Practice

Transfer

• Approved copy: keep it in a document; paste it for edits, never regenerate.
• Code you shipped: the diff is the truth, not the chat that produced it.
• Brainstorming: ask for ten variations — variance is the tool.

Next

Foundations complete — eight lessons. Next wave: Prompting, from clear instructions to structured output; follow the curriculum overview to see what lands next. The temperature glossary entry explains the randomness dial.