How to give an AI assistant the context it needs

You asked ChatGPT to “write a follow-up email about the project,” and it gave you something bland and vaguely wrong. The tone was off. It invented a deadline nobody agreed. You spent longer fixing it than you would have spent writing the thing yourself. So you quietly concluded that these tools are overrated.

They’re not. You just asked a stranger to finish a sentence you never started.

The blank prompt is the problem

An AI assistant knows an enormous amount about language in general and nothing about your situation in particular. It hasn’t seen the email thread. It doesn’t know your colleague is prickly about being cc’d, or that the “project” is three weeks late, or that you want to sound firm without sounding annoyed. When you leave those things out, the assistant fills the gaps with plausible guesses. Plausible guesses are exactly what get you a confident, tidy, useless answer.

The fix is not a clever phrase or a magic prompt you found online. It’s context. Give the assistant the same briefing you’d give a sharp new colleague who’s covering for you while you’re out. You wouldn’t say “write the email” and walk away. You’d hand them the thread and explain what you’re trying to achieve.

What “context” actually means

Context is five ordinary things. The material: the actual email, document, notes, or numbers you’re working from. Paste them in. The goal: what you want to happen after someone reads this. The audience: who it’s for and how they’ll take it. The constraints: length, tone, things to avoid, the deadline that’s real versus the one you’d prefer. And, when you can, an example: a past email you were happy with, so the assistant can match your voice rather than inventing one.

Most disappointing results are missing at least three of those. You don’t need all five every time. But the more the assistant can see, the less it has to invent.

A pattern you can reuse

Here’s the shape of a good request. Paste the thing. Say what you want. Say who it’s for. Read the answer critically. Then iterate.

That last step matters more than people expect. The first reply is a draft, not a verdict. Treat it like a rough cut from a competent assistant who’s never met your reader. Read it and react. “Warmer.” “Too long, cut it by half.” “You’ve assumed we agreed a date — we didn’t, take that out.” A second pass with real feedback almost always beats starting over.

Redrafting a work email

Say you need to chase a supplier who’s gone quiet. The weak version is “write a chasing email.” The strong version pastes your last message to them, then says something like: This is the third time I’ve asked. I want the parts by Friday, but I still need to work with these people next year, so keep it civil and short. I’m writing to their account manager, not the boss. Here’s how I usually sign off.

The difference is night and day, because now the assistant knows the stakes. It knows civil-but-firm, not scorched-earth. It knows Friday is the ask. It knows the relationship continues. You’ve turned a blank request into a brief.

Summarising a long document

Long documents are where people expect the most and brief the least. “Summarise this” gives you a shorter version of everything, weighted by nothing. That’s rarely what you need.

Tell it why you’re reading. I’m about to sit in a meeting where this contract will be discussed. I care about payment terms and anything that limits our liability. Ignore the boilerplate. Give me the three things most likely to trip us up. Now the summary has a job. And here’s a habit worth keeping: for anything that matters, ask the assistant to point you to where in the document each claim comes from, then check those spots yourself. A summary is a map, not the territory.

Planning something

Planning a weekend away, a small event, a house move — the assistant is genuinely handy here, but only if it knows your reality. Budget. Dates. Who’s coming. What you hate. “Plan a weekend in Edinburgh” gets you a generic listicle. “Two adults, one who can’t walk far, tight budget, we’ve been before so nothing obvious, arriving Friday night” gets you something you might actually use. Constraints don’t limit the assistant. They aim it.

When it’s confidently wrong

This is the part to take seriously. An AI assistant does not know when it’s guessing. It states invented facts in the same calm, fluent voice it uses for true ones. There’s no wobble, no “I think,” no tell. That fluency is exactly why people get caught out.

So build a little suspicion into your routine. Be most alert when the answer contains specifics you didn’t provide: names, dates, figures, quotes, citations, legal or medical claims. Those are the things it’s most likely to fabricate and the things you can most easily check. If the assistant references a source, look at the source. If it cites a rule or a number, verify it somewhere real before you rely on it. And when the reply sounds too smooth for a messy question, that smoothness is a reason to slow down, not relax.

None of this is technical. You’re not learning to code or memorising commands. You’re learning to brief well and to read the result like an editor — which, as it happens, is a skill you already use every time you hand work to another person. The assistant is just faster, tireless, and in need of a clearer briefing than any human would ever tolerate. Give it the context. Then check its work.