How to turn a description of a problem into an AI task you can check

You open a chat with AI and freeze in front of the empty line. You have something to ask, but first you want to phrase the request "properly": give the model a role, describe the result, remember the tips from a saved collection of prompts – ready-made requests to AI. It becomes one more job to do before the real one.

Notes have piled up, a difficult conversation is coming, you want to get a grip on a project. You can describe the situation, but you don't yet know what task will come out of it.

Start by telling the story in your own words. Ask AI to restate the task and check whether its understanding matches yours. That is how a good task definition can emerge.

You don't need a final solution in your first message. But as the conversation goes on, you will have to clarify what you want and what result you are prepared to accept.

First, say what is going on

A typical situation: you have a project meeting in two days, and the information about it is scattered across notes. Decisions are written down in one place, your thoughts in another, unanswered questions in a third. You need to prepare for the conversation, but it is not clear where to begin.

You could write "Help me prepare for a meeting" – it is a start, but the model has almost nothing to work with. Or you could say more:

I have a project meeting in two days. From previous conversations I have a few notes, and they mix agreements, my ideas and questions. I want to understand what needs to be discussed so that we don't walk away with different understandings again. I'm not attaching the notes yet – first help me work out what to pull together from them.

This already contains the reason for asking, shows where the difficulty lies and what the person intends to do next. The form of the result – a list of questions, a summary or an agenda – has not been chosen yet, and that is fine: there is enough material to start.

Details that seem obvious to you are especially useful. That you were the only one keeping notes. That some phrases were written down from memory. That one decision has to be made at the meeting and the rest can wait. The model doesn't know any of this until you tell it, and a nicely formatted request will not make up for a missing detail.

Start a big task with a "brain dump"

For a new project or a big task, a technique you might call a "brain dump" helps. Open a chat or an agent – Codex, for example – and talk through everything connected with the task: the idea, where it came from, your doubts. As you go, you can return to the beginning and discover that part of the idea is not yet clear even to you.

A rambling account is normal at this stage. If you immediately turn it into a neat task, it is easy to lose a detail that later proves important. After you have finished, ask the model not to do anything yet and to explain how it understood you. Read it, correct inaccuracies, add what you forgot – sometimes it takes several rounds.

Voice is only one option. You can write a text, use your notes or start with a couple of sentences. What matters more is keeping the meaning you want to convey and checking what came out of your explanation.

Ask AI to restate the task

After a long explanation, it is tempting to jump straight to the solution: the model is already proposing a plan, and all that is left is to pick an item.

But first, pause on how it has understood you. In the meeting example, the restatement might look like this:

You want to compile a list of questions for the upcoming meeting from your notes. Recorded agreements need to be separated from your proposals, and places where confirmation is missing should be flagged. The outcome of the meeting is not yet known – right now you are preparing material for discussion.

A restatement like this is easy to compare with your intention: you can see what the model plans to do and how it interprets the notes. Here is another possible answer:

You need a project management system with a single repository of agreements, owners and automatic reminders.

Perhaps such a system will be useful one day. But between preparing for one meeting and rebuilding how the project is run, several decisions have appeared that you never made. Miss this leap, and you can spend a long time discussing tools and database structure while the original task stays where it was. That is why asking for a restatement is useful even before a detailed plan: it shows the moment your account turned into someone else's version of the task.

Correct the specific discrepancy

A smooth answer is easily mistaken for an accurate one: everything connects, the wording is confident, each point looks reasonable. Check the details.

You wanted to prepare questions, and the model drew up a final project plan.

You mentioned a possible deadline – in the restatement it became an agreed date.

You described your idea – the answer calls it a team decision.

A general "you got it wrong" signals a problem but does little to fix it. Point to the place where the meaning shifted:

We are not choosing a project management system right now – I only need material for the next meeting. And we have not agreed a launch date: in the notes it is my proposal. Rebuild your understanding taking these two corrections into account.

Then read the restatement again: a corrected sentence does not mean the model has revised every conclusion it drew from it. This kind of check also sharpens your own thinking. While reading, you may find that you want to solve several tasks at once, or that you cannot yet explain what decision should come out of the meeting. That is also a useful result: now you know what remains unclear.

Find the question that changes the next step

Not every uncertainty needs to be resolved straight away. When preparing for the meeting, it matters who will read the final document – only you or all the participants: that determines how much context to explain and which personal notes to remove. The colour of a future table, however, is unlikely to help you decide what to do with the notes.

Ask AI to pick out the questions whose answers determine the next action. If it returns a long questionnaire, ask which question to start with and why. In our example, it might be: "What decision do you want to make at the meeting?" If you don't have an answer, say so – then the first task becomes finding the unresolved questions in the notes. "I don't know yet" reflects the real state of the work and is more useful than a random answer given just to fill in every item.

There is no need to keep polishing the description forever either. Once it is clear what small result you need and how you will check it, give it a try.

An agreed description gives you the first action

After a few clarifications, the task might look like this:

From my notes, prepare a draft list of questions for one meeting. Show explicitly recorded agreements and my proposals separately. If the status of an entry is unclear, flag it. Do not invent new deadlines or commitments.

Now you can choose a limited first step: work through one fragment of the notes and see whether the result fits. Use material that you are allowed to share with the chosen service; if it contains unnecessary personal details or confidential information, prepare an anonymised fragment. Then check: has an important question disappeared, has a proposal turned into an agreement, is it clear from the result what you will discuss?

This trial tests something different – how well the model carries out the agreed task. An accurate restatement does not yet guarantee a correct result. If AI has access to files or applications, name the permitted action separately: for example, prepare a draft in the reply without changing the original notes. Publishing the document or sending it to participants is the next, separate decision.

You will still need precise instructions

You can start in plain language, but the quality of your explanation still affects the work. The model needs the source information and an understanding of the constraints and the form of the result, and you need a way to check the answer. The difference is that you don't have to gather these conditions on your own before the first message: you can work them out as the dialogue goes on.

In the meeting example, the precise instructions appeared gradually – work with one meeting, distinguish proposals from agreements, flag what is unknown. They grew out of specific misunderstandings. When an approach that works has repeated several times, save the description as your own template: it will contain conditions whose meaning you understand.

A ready-made prompt from someone else's collection can also help, but you will have to match it to your situation. A template that requires an owner and a deadline for every item may fit poorly with notes in which those decisions have not yet been made.

Practice: work through one task before you start

Take a task you keep putting off because it is hard to explain in full – small enough that you can check the first result. Say what is happening and where the difficulty lies, what you have already tried, what result you need (or that you don't know yet). Ask the model to restate its understanding and compare it with what you meant:

• whether the original goal has been kept;
• whether the circumstances that affect the work have been taken into account;
• whether your decisions are separated from the model's assumptions;
• whether unknown questions have remained unknown;
• whether the task has expanded without your decision;
• whether it is clear what first result you will be able to check;
• whether it is clear which actions are currently allowed.

Correct the discrepancies and choose one step. For a meeting – working through one fragment of the notes; for a text – one section; for a calculation – checking the input data before computing. After the trial, it will be easier to decide what to clarify and whether to continue in the same way.

A prompt to start the conversation

If it is still hard to begin in front of the empty line, use this request. It organises the discussion, but reading the answers and correcting the meaning will be up to you.

Help me work through a task that I can't yet
phrase precisely.

First, I'll explain in my own words what is going on:

[Describe the situation, it can be rambling: what you want to change,
where you get stuck, what you have already tried. If the result or
the constraints are not yet clear, say so.]

In the first step, do not carry out the task and do not change
any files or data.

First, restate how you understood my account:
— what is happening now;
— what I want to achieve;
— which circumstances and constraints matter.

Mark your own assumptions separately. Do not turn
my ideas into decisions already made. If you don't know something,
keep it as an open question.

Pick out the questions whose answers are needed to choose
the next action. Start with the most important one.

Wait for my corrections and then update the task description.
When I confirm that the description is correct, propose
one small first step: what result we will get
and how I will be able to check it.

Move on to execution only after
we have agreed on that step.

You can change the wording. If the task is already clear, there is no need to go through the whole conversation again. And if the description breaks down into several goals, choose one first. The criterion is simple: when you read the final description, you recognise your intention in it and know where to start.

This is also how a conversation with us starts

To discuss a task with Simplit, you don't need a finished technical specification or to know the names of tools. Tell us how the work is organised now, what is blocking the next step and what you have already tried to change. If all you have is a feeling that "something here could be made easier", give one specific example. A good task can perfectly well begin with the words "I don't know yet how to put this" – precision comes in conversation.

Read next:

• to find a task for an AI agent in everyday work – examine your working day before building an AI agent;
• to avoid losing the result and carry on working – where to put the result of an AI conversation;
• if for now you only need AI for yourself – start with what you are already trying for yourself.

You can describe your situation in plain words, doubts included, at a free 30-minute meeting with Simplit. Discuss your task