Why the first request is usually too thin
Most disappointing AI output traces back to a request that left out something the writer knew but the assistant did not. When you ask a colleague for “a quick update,” they know your team, your last conversation and what your manager likes. An AI assistant knows none of that unless you say it.
The assistant is not reading your mind, and it is not searching your files unless you have connected it to them. It predicts a plausible response from the words you gave it. Vague words tend to produce a generic, middle-of-the-road answer. This is a feature of how language models work, not a sign that you are doing something wrong.
The fix is a habit rather than a trick. Before you send a request, check that it states three things: the task, the context and the format.
Step one: state the task
The task is the verb at the center of your request. Be specific about what you want done to what.
Weak: “Help with my report.”
Better: “Rewrite the summary section of my report so a non-specialist can follow it.”
Useful task verbs include summarize, rewrite, compare, list, explain, outline, critique and draft. If you want more than one thing, say so in order. A request that mixes three tasks often returns three half-finished answers; splitting it into steps usually works better.
It also helps to say what the task is not. “Suggest improvements, but do not rewrite the whole thing” prevents an overeager full rewrite.
Step two: add the context that changes the answer
Context is the background that would change what a good answer looks like. Ask yourself what a smart new hire would need to know to do this well. Typical items are:
- Audience: who will read the result, and how much they already know.
- Purpose: what the result is for, such as informing, persuading or recording a decision.
- Constraints: a word limit, a deadline, things to avoid, or a tone to match.
- Source material: the text, notes or data the answer should be based on.
Include only what matters. A long dump of background can bury the part that counts. And be careful about what you paste: names, account details and confidential material do not always belong in a request, a point covered in more depth under prompt context in our guide on sharing background without oversharing. Check your organization’s AI use policy if you are unsure what is allowed.
Step three: name the format
If you do not say what shape the answer should take, the assistant will pick one, and it may not be the one you can use. Say whether you want:
- a short paragraph, a numbered list, or a table;
- a certain length, such as “three sentences” or “under 150 words”;
- headings or no headings;
- a specific order, such as the recommendation first and the reasoning after.
Format instructions are cheap and make a large difference. “Give me a table with three columns: option, benefit, risk” is far easier to use than a wall of text you then have to reorganize.
Putting the three together
Here is a before-and-after example for a common workplace request.
Before: “Write something about our new schedule.”
After: “Draft a short announcement for our warehouse staff about a change to shift start times. The audience is hourly employees who read on their phones. The tone should be friendly and direct. Use three short paragraphs: what is changing, when it starts, and who to ask with questions. Leave the dates as [DATE] placeholders.”
The second version takes about thirty seconds longer to write and usually saves several rounds of correction. Notice that it uses placeholders for details you would rather fill in yourself. That is a handy habit for anything you do not want the assistant to guess.
Iterate instead of starting over
Your first message does not have to be perfect. Treat the exchange as a conversation. If the answer is close, say what to change: “Shorter, and remove the second paragraph,” or “Keep the structure but make the tone more formal.” Specific feedback works better than “try again.”
If the conversation drifts or the answers get worse, starting a fresh request with a cleaner, combined instruction is often faster than patching the old one.
Ask for what you need to check
An assistant can sound confident while being wrong. These errors, sometimes described as fabricated answers, are more likely when you ask for specific facts, figures or sources that the tool cannot actually look up. You can reduce the risk by asking it to say when it is unsure, to separate what came from your material and what it added, and to avoid inventing details. None of this removes the need to review the result, which is why a routine for fact checking matters for anything you will share or act on.
A short checklist before you press send
- Does the request name one clear task?
- Have I said who the audience is and what the result is for?
- Have I included the material the answer should rely on, and left out what is private or unnecessary?
- Have I described the format and length?
- Is there anything I want the assistant to avoid or leave as a placeholder?
If you can answer yes to these, you are already ahead of most first attempts. Over time you will notice which phrasings work for your own tasks. Saving them is the natural next step, and a prompt library makes those habits easy to share.
Results will vary between tools and between runs, and no wording guarantees a perfect answer. The goal is simply to give the assistant enough to work with and to keep a person responsible for the final result.