> For the complete documentation index, see [llms.txt](https://docs.artific.nl/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.artific.nl/en/for-everyone/getting-better-results.md).

# Getting better results

How to ask, so you get something useful back.

None of this is about clever tricks. It is about giving the assistant what it needs, the way you would brief a capable colleague who has just joined.

## Say what you actually want

The most common reason for a disappointing answer is a question that did not say enough.

| Instead of                                  | Try                                                                                                                                                        |
| ------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------- |
| *"Write something about our leave policy."* | *"Write a short internal announcement, about 150 words, explaining that unused leave now expires at the end of March. Friendly but clear. For all staff."* |
| *"Summarise this."*                         | *"Summarise this in five bullet points for someone who has not read it, focusing on the financial risks."*                                                 |
| *"Is this allowed?"*                        | *"Under our procurement policy, may a team lead approve a €4,000 purchase without a second signature?"*                                                    |

Four things do most of the work: **what** you want, **who** it is for, **how long** it should be, and **what tone**.

## Give it the context it cannot see

The assistant cannot see your screen, your inbox or your files. It knows what your organisation gave it, and what you tell it.

If the answer depends on something specific (which department, which product, which customer, which date), say so.

## Ask, then refine

Treat the first answer as a draft. This is the single biggest difference between people who find these tools useful and people who do not.

> *"Draft a reply to this complaint."* *"Too formal. More human, and acknowledge the delay properly."* *"Good. Add that we will call them on Thursday."*

Three rounds took less time than perfecting one question, and got a better result.

Say what is wrong rather than restating everything: *"too long"*, *"you have missed the deadline"*, *"less corporate"*.

## Ask it to think before it answers

For anything with reasoning in it, ask for the working:

> *"Compare these two options against our requirements and explain the trade-offs before recommending one."*

An assistant asked to explain its reasoning gives better answers than one asked only for a conclusion, and you can see where you disagree.

## Give it an example

If you want output in a particular shape, show it one:

> *"Summarise each of these in this format: Client, Issue, Action, Owner."*

An example is worth several paragraphs of description.

## Start a new chat when you change subject

A conversation carries everything said so far. That helps on follow-ups and hurts when you move on: the old context pulls answers off course. **New chat** for a new topic.

## Check the answer

* **Look at the sources.** Where an answer shows **Sources used**, that is the fastest way to verify it.
* **Be sceptical of specifics.** Numbers, dates, names and amounts are where confident errors do the most damage.
* **Notice when it does not say it does not know.** An assistant with no information will sometimes produce a plausible answer anyway.

## Use the right feature

| If you…                            | Use                                                                   |
| ---------------------------------- | --------------------------------------------------------------------- |
| Have a one-off question            | An [assistant](/en/for-everyone/assistants/chatting.md)               |
| Do the same task repeatedly        | An [AI Toolbox element](/en/for-everyone/toolbox.md)                  |
| Are working on one thing over days | A [project](/en/for-everyone/projects.md)                             |
| Keep repeating the same preference | [Personal instructions](/en/for-everyone/assistants/personalising.md) |

## Things that do not work

* **Being polite or forceful.** Neither changes the quality of the answer. Being *specific* does.
* **Asking it to be sure.** It cannot verify itself. Check the sources instead.
* **Very long questions with several unrelated requests.** Split them.
* **Assuming it remembers yesterday.** It does not, unless you have set up [memory](/en/for-everyone/assistants/personalising.md).