lupy.io
Sections
EN
Download

How to write a request AI agents can finish

A good request to an AI team is short, but it names the outcome, the limits and the person it is for. A practical guide with five everyday examples.

Published: 7 min read

People tend to write one of two kinds of requests to an AI. The first is a single vague line: "make me a website." The second is a wall of text that tries to specify everything and ends up contradicting itself. Neither gives a team of agents much to work with.

A request that gets finished sits between the two. It is short, but it carries four things. This post goes through them and shows five requests that follow the pattern.

What happens to your request

It helps to know what the other side does with your words. In lupy.io, the request goes to the manager, which makes a plan: it breaks the big job into small, clear steps and decides what comes first. Each step becomes a ticket with an owner and a status. Specialist agents pick up the tickets.

So your sentence is not an answer to be polished. It is the source material for a plan. The clearer it is about the goal, the better the plan.

Why short but clear

The side that makes the plan reads what you wrote, not what you had in mind. Something that looks obvious to you is not in the plan if it was not written. On the other hand, writing every detail also drowns the plan; the team cannot pick which one matters. So a good request says little, but says the right things.

A practical way: after writing the request, read it aloud once and ask yourself what someone who knows nothing would do with these sentences. If a question comes to mind, add its answer to the request.

The four things to include

You do not need a paragraph for each. One sentence can carry two of them.

Five requests, before and after

Vague: "Do something about our social media." Better: "Draft a weekly Instagram plan for my café, friendly tone, no more than four posts a week, and show me the first week before anything is scheduled." The outcome is a plan, the audience is implied by the café, the limit is four posts, and the checkpoint is the first week.

Vague: "Look into competitors." Better: "Compare my rivals' prices for the same five products and give me one page I can read in two minutes." The outcome is a one-page comparison; the limit is the five products and the length.

Vague: "Fix my app." Better: "Find and fix the bug in my app where the cart empties after login, and tell me what you changed." The outcome is a fix plus an explanation.

Vague: "Handle my email." Better: "Put important mail first and sum up this week's email. Do not send any replies." The limit here is the most useful part: it says what must not happen.

Vague: "Report for the boss." Better: "Prepare the monthly revenue report from my sales spreadsheet and chart the three biggest changes." The source and the output are both named.

How to use the examples

You do not have to copy the examples above as they are. Take the pattern: result, who it is for, limit, checkpoint. Fill it in with the words of your own job. In the first few tries, writing all four may feel like effort; after a few requests it comes by itself.

If you cannot describe the result of a job in two sentences, the problem is probably not in how the request is written but in the job not being clear yet. Then first ask yourself: when this job is finished, what should I have in my hands?

Say what must not happen

Notice the "do not send any replies" above. Telling the team what is off limits is as useful as telling it what to do. lupy.io already asks for your OK before it spends money, publishes, messages outsiders or deletes, but a plain-language limit in the request keeps the plan itself away from those steps.

Big jobs are fine, but name the first piece

You can hand over a large job, such as an app plus a website plus a social plan. The team will split it. What helps is saying which piece you care about most, or which must be done first, so the plan puts it at the front. If you have no preference, say so; the plan will choose an order.

When the result is not right

Sometimes the first result misses. That is normal. Read what came back, then say what is wrong in the same plain way: "The tone is too formal, and the second post repeats the first." You do not have to rewrite the original request. Adding a correction is usually enough, and the ticket carries on from there.

If the same thing goes wrong twice, the cause is often missing context rather than a weak agent. Add the missing fact, such as who the audience is or what the budget is, and try again.

Length: how much is enough

Two or three sentences is usually the sweet spot. A single sentence works when the job is obvious. More than a short paragraph is a sign that you are describing several jobs at once, and it is cleaner to send them as separate requests or to list them as separate lines. The team can handle several jobs in one project; it just needs to know where one ends and the next begins.

If you are unsure what the team can take on, the home page lists examples: an ordering app for a neighborhood café, a redone hotel website, a simple puzzle game, a rival price comparison, a week of Instagram posts, a chart of a sales spreadsheet, a summary of the week's email. Use them as a feel for the right size.

A quick checklist

If you can answer yes to most of those, the request is ready. If you want to see how a request moves through the team afterwards, A chatbot or a team of AI agents? follows one from sentence to tickets, and Letting AI agents work without losing control explains the approval step.

The home page also shows a request moving through the team, one step at a time, in the order described here.

In short

A good request says four things: the result, who it is for, the limits, and the point where you want to approve. It does not need to be long; it needs to be clear. Writing what must not happen also pulls the plan in the right direction from the start. If the result does not fit, adding the correction in the same plain language is often enough, instead of rewriting the request from scratch. If the same mistake repeats, look for the missing context. After a few requests this pattern comes by itself, and the results you get from the team become clearly more accurate. Writing a short request is a skill, and like any skill it settles in after a few tries; do not aim for perfect on the first request, write something that can be finished first and adjust by reading the result.