How to write an AI request that works the first time

Most disappointing AI answers are not really the AI’s fault. They come from a request that left the important part in your head. The good news is that there is no secret wording to learn. There is one habit, and it takes about thirty seconds before you press send.

Brief it like a capable new colleague

The best way to talk to an AI agent is the way you would brief a smart, willing colleague in their first week. They are quick and capable. But they have not been in your meetings, they do not know your clients, and they cannot read your mind.

That one picture explains most of what goes wrong. When an answer misses, it is usually because the agent had to guess something you knew and did not say. A new colleague would have guessed too.

Imagine handing a new starter this note: “Can you sort out the Henderson thing?” They would have to come back with five questions. Now imagine this one instead.

Try saying:

Henderson Builders say we overcharged them on invoice 4471. Read the invoice and their email (attached), check the rates against our current price list, and tell me whether they are right. If they are, draft a short, friendly reply offering a credit.

Same job. Nothing left to guess. That is the whole skill, and the rest of this article is about doing it quickly.

A good test before you send: could a smart stranger do this from what I wrote? If not, add what they would ask you.

The five ingredients of a good request

Most good requests contain the same five things. You will not need all five every time, but checking for them is the fastest way to improve an answer before you even get it.

1. The goal

What you want, and what it is for. “A summary for the board” and “a summary for me” are different jobs. The purpose tells the agent what to keep, what to drop and how much detail matters.

2. The context

Who it is for, what has happened so far, which files matter, and what you have already tried. The agent knows nothing you do not tell it or point it to. If last month’s figures were wrong, or the client is touchy about price, say so.

3. What done looks like

The shape of the answer: length, format, tone, and where to save it. “Three bullet points.” “A table with these columns.” “A one-page Word document.” “An email I can send as is.” Without this, the agent guesses, and guesses vary.

4. An example

One good example of what you want beats a paragraph describing it. If you want an email that sounds like you, paste one you sent before. If you want a report in your house style, point to an old one. Describing a style leaves room to guess. Showing it takes one line.

5. The limits

What not to do, what not to change, and what to ask you about first. “Do not email anyone.” “Keep the current prices.” “Ask before deleting anything.” This is the ingredient people forget most. If something would be bad to get wrong, such as emailing a client or changing a price, say so.

A weak request, made strong

Here is a request most of us have sent at some point:

“Write a summary of the staff survey.”

The agent will produce something. But is it for the board or for the staff? One page or ten? Should it include the angry comments? Should it name people? It has to guess each of these, and every guess is a chance to miss.

Now the same request with the five ingredients in it.

Try saying:

Summarise the staff survey results (attached) for the leadership meeting on Thursday. They have ten minutes to read it, so one page at most. Start with the three things that changed most since last year, then the top complaints with a typical quote for each. Do not name anyone or quote anything that could identify a person. Use the same layout as last year’s summary, which is also attached. Save it as a Word document.

Look at what each part does:

• Goal: a summary for the leadership meeting.
• Context: where the results are, who reads it, how long they have.
• Done looks like: one page, a set order, a Word document.
• Limits: no names, nothing that identifies a person.
• Example: last year’s summary shows the layout.

It took an extra minute to write. It will probably save you three rounds of corrections.

Ask for the plan first on big jobs

On a big job, a wrong turn early on can waste a lot of work. Fixing a plan takes one message. Fixing a finished report built on the wrong idea takes many. So for anything large, ask to see the plan before the agent starts.

Try saying:

Before you start, show me your plan: the steps, which files you will use, and what you will hand me at the end. Wait for my OK.

When the plan comes back, check four things:

• Did it understand the goal, or a slightly different one?
• Is it using the right files and sources?
• Is anything missing, or is it doing more than you asked?
• Is the finished thing what you actually need?

Say you ask for a review of twelve supplier contracts. The plan says it will compare prices. Reading it, you realise you care more about notice periods and liability. You reply in one line, the plan changes, and no work is wasted.

For a plan that really matters, get a second view before work starts. In Interweave OS you can ask the agent to have a helper from another model family check the plan on its own, because different families tend to catch different mistakes.

Small jobs do not need a plan. Asking for one on a two-minute task just slows you down.

When you are not sure what you want yet

You do not need to know exactly what you want before you start. Often you only find out by talking it through. Just tell the agent that is what you are doing.

Try saying:

I want to improve how we onboard new clients, but I’m not sure where to start. Ask me up to five questions, one at a time, then suggest two or three approaches.

“One at a time” keeps it a conversation instead of a form to fill in.

Rambling is fine, too. A few paragraphs of “this is what’s going on, this is what worries me, this is what I tried” often give the agent far more useful context than a short, tidy prompt. It can sort the mess. It cannot invent the context you left out. Dictation helps here: talking is faster than typing, so you say more.

On an important job, you can also ask what it needs before it begins.

Try saying:

Before you start: what would you need to know to do this really well? Ask me.

When the first answer misses

Sometimes the first answer will miss. That is normal, and quick to fix if you say what is wrong and why.

• Weak: “No, try again.”
• Strong: “Too formal. These are long-time clients and we’re on first-name terms. Keep it to five lines and drop the sales pitch at the end.”

The weak reply gives the agent nothing to go on, so the second try is another guess. The strong one says what to change and why, so it can get the rest right too.

Say what worked as well, or a rewrite may throw out the good parts. And if you cannot see why it went wrong, ask: “What were you assuming?” The answer often reveals a wrong belief you can correct in one line.

Resist fixing the same kind of thing by hand. A final touch is fine. But if you quietly fix it yourself, the agent never learns what you wanted in that chat, and the next draft has the same problem.

Say it once, not every time

Some guidance you give again and again: Australian spelling, short answers, our clients are law firms, never promise a delivery date. Typing it into every request wastes time, and the day you forget, the answer misses.

Most AI tools have somewhere to put standing instructions. Use it. A simple rule: if you have said the same thing three times, it belongs there. If it applies to one kind of job, such as your monthly report format, it belongs in a reusable skill or template instead. Interweave OS has both: layered instructions for you, your team and the organisation, and skills the agent picks up by itself whenever a request matches.

Do not overdo it

A quick question needs a quick request. “What is the GST on $4,200?” is fine as it is. Save the five ingredients for when the job matters, or when the first answer missed.

How Interweave OS does it

Interweave OS is built around briefing agents well. Every chat starts with your organisation’s, your team’s and your own instructions already loaded, so the context you would otherwise retype is there before you write a word. Skills turn a request you have perfected into a one-liner the whole team can use. And when you want a plan checked, the agent can hand it to a helper from another model family in the same chat.