---
id: KB-MA-014
url: https://app.codecontract.io/help/marta/the-first-five-questions
idioma: en
categoria: marta
audiencia: usuario
nivel: basico
actualizado: 2026-08-13
tambienEn: [es]
relacionados: [KB-MA-002, KB-PS-011, KB-MA-010]
citadoPor: [KB-MA-006]
---

# The first five questions

_What to ask on day one to find out whether this will help you._

**Responde a:** what should i ask the ai on day one · first questions for the assistant · how to start using marta · examples of useful questions

The first time someone sits in front of the assistant they usually ask something generic, get a generic answer, and conclude it is useless. These five questions are chosen for the opposite: they are specific, verifiable, and show what it can do.

## The five

1. **"What expires in the next thirty days?"** — Useful from minute one and checkable at a glance.
2. **"What is waiting on a third party?"** — It shows where work is stuck, which is rarely where you thought.
3. **"Summarise what has happened with [a specific supplier]"** — Pick one you know well: then you can judge whether the summary is good.
4. **"How long do we take on average to close a file?"** — A figure that normally costs half a morning to produce.
5. **"What documents exist for [an identifier]?"** — A tax ID, a registration, an order number. Proof that it finds by content.

> [!IMPORTANT]
> The third tells you best whether you can rely on it. Asking about something you know by heart, you will see in two seconds whether the summary is right, whether it omits something important, or whether it treats a wrong value as good.

## What not to ask on day one

| Question | Why it disappoints |
| --- | --- |
| "What can you do?" | It answers in the abstract and shows nothing of yours |
| Something in a file you cannot see | It will say there is no record, and rightly |
| A question with four conditions | Two simple questions work better |
| Something you uploaded a second ago | It may take a moment to become available |

> [!WARNING]
> And do not ask about scanned documents that were never read, expecting answers about their content: they exist, they display, but they say nothing about themselves. If your archive is mostly unprocessed old scans, that shows up right here.

## What to look for in its answers

**En corto**

- Whether it says where each value comes from.
- Whether it distinguishes what is recorded from what it infers.
- And whether asking the same thing twice gives the same answer.

That last point matters more than it seems: an answer that changes between attempts signals it is at the edge of what it can answer with what exists.

> [!NOTE]
> Every question consumes, including those that yield nothing. It is not much, but worth knowing before spending an afternoon trying phrasings.

**Can I ask in another language?**

Yes, and it answers in the one you use.

**Can it prepare things?**

Yes: drafts, lists, processes. Preparing is different from executing.

**Does it see what my colleagues see?**

It sees what you see, no more and no less.

## Ejemplos

**Someone tries the assistant by asking what it can do and dismisses it.**

- Tries the five specific questions
- Checks the summary of a supplier they know well

→ Finds in five minutes two figures that would have cost a morning.

**The first question is far too broad.**

- Starts with a specific file

→ The first answer is already useful.

**It is tried with a question that has no answer.**

- Starts with what is clearly in the documents

→ The first impression is the right one.

**It is dismissed after a poor answer.**

- Reformulates more specifically

→ The second answer changes the view.

**The question omits the period.**

- Adds the date or the range

→ The answer points at what was sought.

**It is tested with made-up data.**

- Tests with a real case

→ The result says something about your work.