# prefer (10-prefer.com)

When people say they prefer something, they mean they like it better than alternatives. It is a soft statement of taste, not a command. A person might prefer tea to coffee, or prefer to meet on Tuesday rather than Monday.

## What a model may hear
- HTTP Prefer header (web standards and REST APIs): treat the word as a request header that asks the server to return partial content, handle asynchronously, or apply specific processing
- preference relation (economics, game theory, and decision theory): model it as a formal ordering over choices, often assuming transitivity and completeness axioms
- CSS prefers-* media query (web design and accessibility): read it as a user-system preference signal like prefers-reduced-motion or prefers-color-scheme
- Reinforcement Learning from Human Preferences (AI training and alignment): interpret it as training signal data comparing outputs pairwise to shape model behavior
- database query optimizer hint (SQL and query planning): understand it as a directive to prefer an index, join order, or execution plan

## Where people and models part ways
- Says: "I prefer shorter answers"
  Means: please be concise when you reply to me
  May be taken as: the model may treat this as a system-level preference to be stored and applied globally, or ignore it as non-binding
  Say instead: "Please keep your answer under three sentences."
- Says: "We prefer to handle this async"
  Means: let us deal with it later, not right now
  May be taken as: the model may generate code using the HTTP Prefer: respond-async header or async/await patterns without confirming
  Say instead: "Let us postpone this. Do not use async headers or patterns unless I ask."
- Says: "I prefer the first option"
  Means: choose A over B
  May be taken as: the model may log this as training preference data or apply it to future unrelated choices
  Say instead: "For this specific choice, select the first option."
- Says: "My users prefer dark mode"
  Means: most users like dark backgrounds
  May be taken as: the model may write CSS using prefers-color-scheme without checking actual browser support or user settings
  Say instead: "Most users want a dark theme. Show me how to detect that choice, not assume it."

## Tips
- Say 'for this request' or 'right now' to keep a preference local and temporary.
- If you mean a web standard, name it explicitly: 'the HTTP Prefer header' or 'prefers-reduced-motion'.
- Ask the model to confirm before it changes settings or remembers your stated preference.
- Distinguish taste from instruction: 'I like' is weaker than 'do this'.
- When giving feedback on model outputs, say 'this answer was better' rather than 'I prefer this' to avoid confusion with RLHF signals.

## Often confused with
- preferable: an adjective judging quality, not a request
- preferred: often a stored default or configuration value
- priority: ranking by urgency, not taste
- option: a possible choice without the liking attached
- bias: a systematic skew, not a personal taste
- favor: similar in speech, rarely a technical header