Four things shape what a caller actually hears. They interact, which is why changing
one and being surprised is so common.
The opening line
The first thing the agent says is set separately from the prompt, and it matters out
of proportion to its length.
On an outbound call it decides whether the person stays on the line. On an inbound
call it sets what they think they are allowed to ask for.
It can include values from the call’s context, such as a name, so the same line reads
differently per contact.
If you change one thing about an underperforming agent, change this. It is one
sentence and it gates everything after it.
The prompt
The prompt is the agent’s standing instructions. It is read before every conversation
and it is where behaviour actually lives.
Some things that hold true regardless of what you are building:
Simple instructions beat complex ones. A short prompt that the agent follows is
worth more than a thorough one it approximates.
Structure helps. A long prompt split into sections is followed more reliably than
the same content as prose.
Examples beat description. Showing the agent a good exchange does more than
describing one.
Write for speech. The agent reads this aloud. Unusual words and proper nouns need
spelling out phonetically, or they come out mangled.
Prompts can be saved and reused across agents. See
Prompts.
The voice
A voice has a provider, and usually a speaking speed. Language is set separately, on
the agent. Voices differ in more than timbre. Some handle interruption and pacing better than others, and the only reliable
test is hearing one read your own opening line rather than a sample.
You can also bring in a voice from a provider you already use. See
Voices.
Skills
Skills are the difference between an agent that talks and an agent that does things.
Each one adds an ability and, usually, some instructions of its own that the agent
follows alongside your prompt.
They cover roughly four jobs:
- Ending the conversation well, by hanging up cleanly or agreeing a callback.
- Getting to the right person, by transferring, handling a screener, or working
through a phone menu.
- Doing something concrete, such as booking an appointment or sending a message.
- Knowing things, by searching content you have stored.
Skills add instructions to the agent, so switching on several at once can crowd your
prompt. If behaviour changes after adding a skill, that is why. Add them one at a
time and test between.
Context
A call can start with values already known, such as the person’s name or an order
number. Those are available to the prompt and to the opening line.
On a campaign they come from your contact list. On a test call you set them by hand.
Personalisation only works when they are actually present, which is why a test call
with empty context sounds worse than the real thing.
Where to go next
