Give useful context
- prompting
- context
Context is everything the model needs to know that is not already in your request: who the result is for, why it matters, what is already known, and what material to work from. Without it, the model fills gaps with the average of everything it has seen — and that average is rarely your situation.
The four slots
- Audience — who reads or uses this. A client, your manager, a beginner, a machine.
- Purpose — what happens next. Send as-is, make a decision, learn something.
- Knowns — facts and decisions already settled, plus constraints that must survive.
- Material — the text, data or documents to work from. Paste them; do not describe them.
Two sentences usually cover audience and purpose; knowns and material are often one line plus a paste. That is the difference between “rewrite my email” and a result you send without edits.
Context is not a dump
More is not always better. Everything you add competes for attention inside the context window — a 40-page dump with one buried instruction gets worse results than ten relevant lines. Add what changes the answer; summarize the rest.
A bad example
Ask: “Rewrite this email in a more professional style.”
Professional for whom? The model aims at a polite generic register — maybe too stiff for the colleague, too warm for legal.
A better example
Ask: “Rewrite the draft below as a reply to a client who has waited a week: apologetic but confident. Keep every commitment we made. Maximum five sentences.”
Audience, situation, tone, boundary, length. The draft can now land on the first try — or fail in ways you can name and fix.
Why it works
Every context sentence rewires the probabilities that follow. The model cannot infer your client from the draft alone — but given the client, the draft transforms. Context is the cheapest quality upgrade in prompting, and it doubles as documentation: to write it, you had to know your audience.
A quick audit for any draft prompt: can you point at the who, the why and the what-you-already-know? A missing “who” produces the wrong tone; a missing “why” produces the wrong depth; missing known facts produce questions you had already answered. Three lines restore all three — and the answers stop being generic.
Practice
Which prompt carries the context?
You are sending a reply to a client who has been waiting a week for an answer about their refund.
Which prompt is more likely to produce a send-ready draft?
Hint
Which prompt tells the model something it could not possibly know?
Why this is the answer
Context sentences are the cheapest quality upgrade in prompting: audience, purpose, knowns and material rewiring every prediction that follows. The client is not in the draft — supply her, and the draft transforms.
Transfer
- Asking for analysis: say what decision it feeds — insight changes shape.
- Delegating writing: name the reader and the register before the topic.
- Sensitive domains: state the constraint (“we cannot promise dates”) — it survives into every draft.
Next
Next: Write clear instructions — one job per instruction, checkable by a stranger. The prompt glossary entry is the one-line version.