What your agent says

What your agent says comes down to two pieces:

1. The opening line. The very first thing it says when someone picks up.
2. The instructions. The brief that guides everything after that.

​

The opening line

The opening is generated from the script’s opening step. Put specific opening words there. There is no separate fixed-opening setting. This is the first sentence out of the agent’s mouth. It sets the tone for the whole call, so keep it short, human, and honest about who’s calling.

• Good: “Hi, this is Sarah from Coleman Roofing, do you have a quick minute?”
• Not so good: “Hello, this is an automated call regarding your recent inquiry submission.”

The first one sounds like a person. The second one sounds like a robot reading a form, and people hang up on forms. You can leave the opening line blank. The agent then composes its own first words from the instructions on every call, which usually sounds more natural and never greets twice.

​

Who speaks first

Open the agent’s Settings tab and the How calls start card. Two choices:

• Greet first. The agent speaks the moment the line opens. The usual choice for a call the agent places.
• Wait for hello. The agent stays quiet until the person speaks, the way a human answers a ringing phone. The usual choice for a call the agent answers. If the person stays quiet, the agent speaks after the number of seconds you set, so nobody is left listening to silence. The dashboard allows one to fifteen seconds. An agent set up before this setting existed may have no limit at all; the card says so, and choosing a number fixes it.

​

The instructions

The instructions are the agent’s brief. This is where you tell it who it is, why it’s calling, and how to handle the conversation. Think of it as onboarding a new hire, not writing a movie script. A good brief usually covers:

• Who the agent is. Its name, your company, its role.
• Why it’s calling. The reason, in one plain sentence.
• The goal. The one thing you want out of the call.
• What to do when. “If they’re interested, do this. If they’re busy, do that.”
• The hard rules. Things it must never do or say.

You don’t need to cover every possible twist. Cover the goal and the common turns, and trust the agent to handle the rest naturally.

​

Editing an existing script

You can ask the builder chatbot to change a sentence, a step, the opening, or the whole script. Name the part you want changed and the result you want. For example: “In the opening, replace ‘Do you have a minute?’ with ‘Is now a good time?’ and keep the rest as it is.” The edit should follow this process:

1. Read the current script. Start from the latest available script, including edits already made in this conversation. Earlier chat messages are not a reliable copy of the current script.
2. Identify the requested scope. A sentence change belongs in that sentence; a request about one step belongs in that step. If the target could mean two different passages, resolve that ambiguity before changing either.
3. Preserve everything outside that scope. Keep unrelated wording, CRM variables, questions and rules. Replacing the whole script is appropriate when you ask for a broad rewrite.
4. Apply the change to the current text. If the passage has changed since it was read, read it again and use the new version. Do not rebuild missing text from memory or force an old edit onto a newer script.
5. Check the result. Confirm the intended wording changed and surrounding content was preserved. The chatbot should report an edit as saved only when saving succeeded; a staged edit or failed save must be described as such.

​

Write for the ear, not the eye

The agent’s words are spoken, not read. So write the way people talk.

• Do say “January fifteenth.” Don’t write “1/15.”
• Do say “call us at four one five…” Don’t paste a raw link or a symbol the agent would have to read out awkwardly.
• Do keep sentences short. Long sentences sound out of breath.
• Do let it use small natural words (“sure,” “no worries”) so it sounds real.

​

Pull the person’s details into the prompt

Start with Variables and context to check the source, availability and missing-value behavior of a supported variable. Your instructions can include blanks that fill themselves in for each call. On a GoHighLevel sub-account you map each blank to a custom field on the contact (their name, their details) or a custom value on the location (your business name, the transfer number). On a sub-account that keeps its contacts here, you define your own fields and read them by key, for example {{budget}}; see Contacts without a CRM. The agent then speaks the real value from that person’s record, not a placeholder. Map the first name field and the agent greets each person by their actual name. It’s like a mail merge, except the data comes live from the contact record on every call.

​

In the API

Everything on this page lives on the agent under config.llm:

• generalPrompt is the instructions.
• beginMessage can appear on reads of older agents. It is not writable: the opening is generated from the script’s opening step.
• startingSpeaker is who speaks first: agent greets first, user waits for the person.
• beginAfterUserSilenceMs is how long a waiting agent gives the person before it speaks first, in milliseconds, from 1000 to 15000. It only applies when startingSpeaker is user, and a waiting agent always has one: switch to user without sending it and the agent gets 3000.

Read them with Get agent and change them with Update agent. hasCustomPrompt on the agent tells you whether it has been given a prompt of its own or still carries the one it was created with. The blanks that fill themselves in are GoHighLevel custom fields and custom values. Their fieldKey values, such as contact.first_name, are listed by GoHighLevel resources. Not sure which of two openers is better? Run a split test and let the leads decide. Starting from nothing? Your first script builds a draft from a few details about your leads and the conversation you want.