> ## Documentation Index
> Fetch the complete documentation index at: https://docs.brilo.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Write what your agent says

> Change your agent instructions in Brilo so it greets callers your way, handles the awkward moments, and never promises what you cannot deliver.

Your agent's instructions decide everything it says on a call. You can rewrite them at any time in
the **Prompt** tab, and changes apply to the next call rather than the one happening now. This is the
highest value edit you will make: five minutes here beats an hour of anything else.

## Before you start

* An agent, even an unfinished one. See [Create your first agent](/agents/create-your-first-agent).
* Ten minutes and knowledge of how you want awkward calls handled.
* Facts like prices and opening hours belong in knowledge instead, where you can change them without
  rewriting. See [What your agent can learn](/knowledge/what-your-agent-can-learn).
* You will be working in your agent's **Prompt** tab.

## Steps

<Steps>
  <Step title="Open the Prompt tab">
    Open your agent and go to **Prompt**. If Brilo built the agent for you, there is already a full
    set of instructions here. Read it before you change anything.
  </Step>

  <Step title="Fix the greeting first">
    The first line callers hear matters more than the rest of the page. Say who is answering and
    invite the caller to speak, for example "Thanks for calling Marston Dental, how can I help?"
    You can also set a fixed greeting in the **Configure** tab.
  </Step>

  <Step title="Say what the agent is there to do">
    List the jobs in order of how often they come up: taking bookings, answering questions about
    services, taking a message. An agent told what matters most handles the common call better.
  </Step>

  <Step title="Say what to do when it cannot help">
    This is the part most people skip and the part that decides whether callers are satisfied. Write
    down what happens when your agent does not know: take a message, offer a callback, or pass the
    call to a person. See [Pass the call to a person](/actions/pass-the-call-to-a-person).
  </Step>

  <Step title="Write the things it must never do">
    Be explicit. Never quote a price for custom work, never promise a same day appointment, never
    give clinical advice. Your agent follows a clear prohibition well and guesses when you leave a
    gap.
  </Step>

  <Step title="Save and test the change">
    Save, then call your agent and try the exact situation you just wrote instructions for. Reading
    the words back is not testing them. See
    [Test your agent yourself](/go-live/test-your-agent-yourself).
  </Step>
</Steps>

## You will know it worked when

* Your agent uses your greeting word for word on a test call.
* It behaves the way you described in the awkward situation you wrote for.
* It declines the thing you told it never to do, even when you push it.

## What to include, in order

Instructions work best in this order, because your agent reads the whole thing every time.

| Part                  | What goes in it                                                        |
| --------------------- | ---------------------------------------------------------------------- |
| Who it is             | The business name, and that it is answering for you                    |
| The greeting          | The exact opening line                                                 |
| What it should do     | The jobs it handles, most common first                                 |
| How to behave         | Tone, how much to say, whether to use the caller's name                |
| When it cannot help   | Take a message, offer a callback, or transfer                          |
| What it must never do | Prices it cannot quote, promises it cannot make, advice it cannot give |
| How to end a call     | The closing line, and what to confirm before hanging up                |

## What else you can do here

* [Set the tone and patience](/agents/give-your-agent-a-personality) in the **Configure** tab rather
  than describing it in words.
* [Add your prices and policies as knowledge](/knowledge/add-a-document-or-website) instead of
  writing them into instructions.
* [Teach it how to say your business name](/knowledge/teach-it-names-and-terms) if it gets that
  wrong.
* [Save a version](/agents/versions) before a big rewrite so you can go back.
* [Copy a working example](/agents/example-prompts) rather than starting from an empty box.

## If it did not work

**Your agent ignores part of your instructions.** Usually the instructions are long and
contradictory. Cut anything that repeats, and put facts into knowledge instead. Long instructions
compete with themselves.

**Your agent says something you never wrote.** It is filling a gap you left. Find the situation it
mishandled and write the rule explicitly. See
[Why it answered wrong](/knowledge/why-it-answered-wrong).

**Your changes did not take effect.** Check the change saved, and remember a call already in progress
finishes on the previous instructions.

Anything else: [My agent sounds wrong](/agents/my-agent-sounds-wrong).

## FAQ

<AccordionGroup>
  <Accordion title="How do I change what my agent says when it answers?">
    Edit the greeting in your agent's **Prompt** tab, or set a fixed custom greeting in the
    **Configure** tab. The next caller hears the new wording, and a call already in progress finishes
    on the old one.
  </Accordion>

  <Accordion title="Do my changes apply to calls that are already happening?">
    No. A call in progress finishes on the instructions it started with, and the next call uses your
    new ones. That means you can safely edit during the working day without interrupting anybody.
  </Accordion>

  <Accordion title="How long should my instructions be?">
    Long enough to cover the greeting, the common jobs, the awkward moments, and the prohibitions.
    Beyond that, extra length works against you: your agent reads all of it every turn and
    contradictory instructions are the main reason an agent seems to ignore you. Put facts in
    knowledge, not here.
  </Accordion>

  <Accordion title="Where do I put opening hours, in instructions or knowledge?">
    Knowledge. Anything that changes independently of behaviour, including hours, prices, and
    policies, belongs there so you can update it without rewriting your agent. Instructions are for
    how it behaves. See [What your agent can learn](/knowledge/what-your-agent-can-learn).
  </Accordion>

  <Accordion title="Can I stop my agent saying it is an AI?">
    You control what it says about itself, and most callers work it out anyway. The agents that go
    down best do not pretend otherwise when asked directly. Some places also require disclosure, so
    check what applies to you before instructing it to deny it.
  </Accordion>

  <Accordion title="Why does my agent ignore part of my instructions?">
    Almost always because two parts of the instructions conflict, or because the page is long enough
    that the important rule is buried. Cut repetition, put facts in knowledge, and state the rule
    that is being ignored once, plainly, near the top.
  </Accordion>

  <Accordion title="How do I make my agent shorter and less chatty?">
    Tell it plainly to keep answers to one or two sentences unless asked for detail, and set the tone
    to a concise option in the **Configure** tab. Then call it and check, because chattiness is the
    thing customers notice fastest.
  </Accordion>

  <Accordion title="Can I give my agent a script it must follow word for word?">
    You can require exact wording for specific moments, like the greeting or a closing line, and that
    works well. A full word for word script for the whole call does not, because callers do not follow
    the script and your agent needs room to answer what was actually asked.
  </Accordion>
</AccordionGroup>

## Related

* [Example instructions](/agents/example-prompts) to copy rather than compose
* [What your agent can learn](/knowledge/what-your-agent-can-learn) for where facts belong
* [Test your agent yourself](/go-live/test-your-agent-yourself) to check every change
* [Pass the call to a person](/actions/pass-the-call-to-a-person) for the fallback you just wrote
* [My agent sounds wrong](/agents/my-agent-sounds-wrong) for pace and pronunciation