What Context to Give AI to Get Good Answers

A carpenter gets a one-star review on Google: a customer complains about a late installation. He asks the AI to "reply to this negative review" and gets an apology that admits a lack of professionalism and offers a goodwill gesture. But the delay came from a supplier, the customer had been warned, and he has no desire to give anything away.

The AI did what it could with what it had. It was missing two lines, and you shouldn't give it twenty either. Choosing well what to tell it takes a bit of method.

The only test you need: would it change the answer?

For each piece of information, ask yourself one question: if it were different, would I want a different answer?

For our carpenter, "the delay came from the supplier and the customer was notified on the 3rd" changes everything. "My shop was founded in 1998" changes nothing, so that stays out. This test avoids both pitfalls: the skeletal prompt and the life story in which the instruction drowns.

Here's the revised request:

Write my public reply to this Google review:

[PASTE THE REVIEW]

Actual situation: I'm an independent carpenter. The installation delay came from a delivery from my supplier. I called the customer two days ahead to let her know, and the installation was done three days later.

Goal: future customers reading this review should see that I stay calm and that I give notice when there's a problem. I'm not offering a goodwill gesture and I don't want to blame the customer.

Write 70 words maximum, in the first person, with no stock phrases like "we sincerely apologize for any inconvenience." Don't add any fact I haven't given you.

The five types of information that matter

Useful context almost always falls into one of these categories. You don't have to fill them all in: apply the test to each.

The "already tried" category is the most neglected. Without it, the AI serves you the first ideas anyone would have, the ones you've already tested.

Technical context: the bug example

The developer's case shows the difference between a description and evidence. Before:

My React app crashes when I click submit. Why?

The AI can only answer with a list of general causes. After:

My React form crashes on submit.

Environment: [FRAMEWORK AND VERSION], [FORM LIBRARY].
Expected behavior: [WHAT SHOULD HAPPEN].
Observed behavior: [WHAT HAPPENS].
Exact error message:
[PASTE THE ERROR]

Relevant code:
[PASTE THE COMPONENT, WITHOUT ANY KEY OR SECRET]

Already checked: [WHAT YOU TRIED].

Give me the two most likely causes, how to confirm each one, then the fix. If you need another file to decide, ask for it before suggesting a fix.

The logic holds outside of code. A teacher who wants to adapt an exercise won't say "my students can't do it," but will paste the problem statement, two typical wrong answers, and what has already been explained in class. Raw facts beat your interpretation, because the AI may see something different than you do.

When context becomes a problem

More context doesn't always mean better answers. Three traps come up again and again.

Context that steers. "I'm sure the problem comes from the new software" pushes the AI to find that result. If you're looking for an opinion, present the facts and keep your conclusion to yourself until you get the answer.

Context that's too long. A whole document pasted for a question about two pages can drown what matters, and every tool has a limit on what it keeps in mind, as our article on tokens and the context window explains. Paste the useful passage when you can. For large volumes, there are methods for working with long documents.

Context that should never have been there. An accountant who wants to draft a difficult letter to a client doesn't need to paste in a name, a SIRET number (the French company registration number), and a balance sheet. "Client A, small construction company, tight cash flow" is generally enough. Before sending, reread what you pasted and look at what you shouldn't give a chatbot. For personal or confidential data, check the tool's terms and your organization's rules.

Let the AI tell you what's missing

You don't always know what to specify, especially on a new subject. Add this instruction at the end of your request when the work matters:

Before answering, list the missing information that would really change your answer, five at most, with how each one would change it. Wait for my answers before writing.

The "how each one would change it" part prevents a long list of pointless questions.

Write your context sheet once and for all

Some information comes back in every prompt: your business, your clients, the way you write. Write it up once, in ten lines, in a separate document. Here's a template:

My business: [JOB, SIZE, AREA].
My usual clients or readers: [DESCRIPTION].
What I offer: [SERVICES OR PRODUCTS].
How I write: [FORMAL OR CASUAL, TONE, WORDS TO AVOID].
What I never do: [PROMISES RULED OUT, TOPICS TO AVOID].

Paste it at the top of your conversations or save it in custom instructions if your tool allows it. Fill in this sheet today, remove what doesn't pass the test, and put nothing in it that you wouldn't put on your home page.