Prompting a Voice Agent on Athena
A starting prompt for an Athena voice agent: who it is, how it listens and yields, and when it passes work on, with the reason behind each part.
An Athena voice agent is really two jobs. One part holds the conversation: it listens, speaks and keeps the person company. The other part does the work: it looks things up, runs tools, follows procedures and finishes tasks, whether that takes two seconds or twenty minutes. Good prompts respect that split. Write the conversation for the voice, and write the procedure for the worker.
Start with what the voice needs to know
A voice prompt should fit on one screen. If it is longer, something in it probably belongs with the worker. Five things earn their place:
- Who it is talking to and why. The business or product, the person on the other end, and what a good outcome looks like.
- How it sounds. Warm or brisk, formal or relaxed, quick or unhurried.
- How it listens. Whether it stays silent while someone talks or lets out the odd "mm-hm".
- How it gives way. What it does the moment someone talks over it.
- What it passes on. The requests it should hand to the worker, and the ones it can simply answer.
Describe behavior and let Athena find the words. Fixed scripts sound fixed, so save exact wording for the few lines that truly require it, such as a legal disclosure.
A template to adapt
Labelled sections make a prompt easy to review and easy to edit later. Fill in the brackets and keep the headings.
Role: You are [name], the voice of [business or product]. You help [who] with [what]. Sound calm and plain-spoken. When someone is frustrated, say so in a few words and go straight to the next useful step.
Listening: Stay mostly quiet while the person talks. A brief sound now and then is fine.
Giving way: If the person starts talking, stop and hear them out.
Passing work on: The worker can [abilities, in everyday words]. Pass a request on when it needs one of those abilities, or when the person changes something already in progress. Answer yourself when the conversation already holds the answer. Never state a result you have not been given.
Tell the voice what can be done, not how
The voice does not need tool names or parameters. It needs a plain list of abilities so it can recognize a request when it hears one: "Appointments: find open times, then book, move or cancel them." For an agent that works on a computer or inside a project, the same idea applies: "Project work: read and change files, run commands and report what happened."
Be concrete about when to pass work on, and include corrections. "Actually, make it Thursday" and "skip that step" both change the job. One thing to watch in testing: saying "let me check that" is not the same as starting the check. Confirm that the work really began.
Keep the procedure with the worker
Everything about getting the task right lives with the worker: how to read a request, which records or files settle a question, the order of steps, what needs approval, what is private, and what to do when something fails. When you move an older agent to Athena, split its prompt along this line. Then read both halves against the original and make sure each rule has exactly one home.
Add rules only when a test asks for them
A prompt grows best from evidence. When a test conversation goes wrong, add one rule, then repeat the same conversation to see that it helped and broke nothing else. Rules that often turn out to be worth having:
- Keep routine answers to a sentence or two, and give one step at a time when someone is following along.
- When a name, date or number is unclear, ask about only that piece.
- Reuse a recent answer unless it may have gone out of date.
- Do not read long content aloud. Summarize it and offer the detail.
Listen to the audio when you judge the result, because timing and tone never show up in a transcript. The testing guide shows how to make those checks repeatable.
