Giving context: bad vs. great results
Context is the force multiplier in prompting. The same task given with thin context and rich context produces results so different they might as well be from different tools.
What context actually does
When you give an AI model context, you're not just adding background — you're constraining the solution space. A model with no context has to guess what "good" means. A model with rich context knows your audience, your constraints, your tone, and your goal. Its output is calibrated to that specific situation rather than an averaged version of all possible situations.
Think of it like briefing a freelancer. A one-line brief produces generic work. A good brief — explaining the project, the audience, the deadline, the previous approaches that didn't work — produces something tailored.
The four layers of context
Not all context is equally valuable. These four types, roughly in order of impact:
- Who the output is for. "My manager" and "a first-year university student" and "a sceptical investor" will change the content, tone, and vocabulary of any response. Always specify the audience.
- What it will be used for. A summary for a quick decision is different from a summary that will be read aloud at a conference. Specify the purpose.
- What constraints apply. Length limits, tone restrictions, things to avoid, terminology to use or not use — constraints sharpen output dramatically.
- What you've already tried. If you say "I've already written a draft but it's too formal", the model knows not to produce another formal draft. This saves multiple rounds of iteration.
How much context is too much?
A common mistake is dumping everything into the prompt hoping something sticks. Relevant context helps. Irrelevant context dilutes the signal and can actually hurt output quality — the model spends attention on noise rather than your actual ask.
A good test: if a piece of context would change what a good response looks like, include it. If removing it wouldn't change anything, cut it. Two to five sentences of context is usually enough for most everyday tasks.
The before/after test. Weak context: "Write an email following up on my meeting." Rich context: "Write a brief, friendly follow-up email to a client after a 30-minute intro call. They seemed interested but hesitant on price. I want to stay warm without being pushy. Keep it under 150 words and don't mention price — just suggest a next step."
Pasting in documents and examples
One of the most underused prompting moves is pasting in real examples. "Write a bio in the style of the following example: [paste]" produces far better results than describing your preferred style in words. You don't need to describe what you want — you can show it.
The same applies to negative examples. "Don't write something like this: [paste bad example]" is a surprisingly effective constraint. The model learns what to avoid just as readily as what to aim for.
Add context layers to a bare prompt
Take this bare prompt and layer in context using the four types above. Aim for 3–5 sentences of context that would genuinely change the output.
Who will read it? What relationship do you have with them? What kind of feedback — on your work, a product, a proposal? What tone fits?
Specify who the recipient is and what they care about. A busy senior colleague gets a different message than a peer who owes you a favour.
What should the message avoid? What format or length is right for this context? What action do you want them to take?
Send the bare prompt, then the contextualised one. Note the difference in specificity, tone, and usefulness of the output.
What to remember
- Context constrains the solution space — the model calibrates its output to exactly your situation.
- The four most valuable context types: audience, purpose, constraints, prior attempts.
- More context is not always better. Include what changes the output; cut what doesn't.
- Pasting in examples — what you want and what you don't want — is often more effective than describing your preferences.