Write the instruction for an agent that talks to customers
Not prompt engineering — job description writing. Scope, tone, refusals, and the few sentences that prevent most embarrassing transcripts.
The field is called Instruction, and the name is a better guide than "prompt" would be. You are not engineering anything. You are writing what you would tell a competent new starter on their first morning: here is the job, here is how we speak to people, here is what you must never do, and here is when to come and get me.
The instruction that works is usually shorter than people expect and more specific. Vague instructions produce vague agents, and every sentence that tries to anticipate a particular question is a sentence that should have been a knowledge base entry instead.
Before you start
- One clear job for this bot.
- Five real customer questions, taken from transcripts rather than imagination.
- A decision about what it must never say.
Steps
Open the bot and find Instruction
It sits in the Identity card under the assistant name and the first message.
Open with what it is and who it serves
"You are a friendly sales assistant for Convarza." One sentence. The agent needs to know whose side it is on before anything else.
State the job as a sequence
"Qualify by budget, team size and timeline, then book a demo." That is a job with an order and an end state, which is far more useful than a list of topics it may discuss.
Write the refusals explicitly, and name them
"Never quote custom pricing — hand off to a human for enterprise." Notice the shape: the refusal and the escape route in the same sentence.
Three or four of these prevent most of the transcripts you would have to apologise for. Typical ones: never promise a delivery date, never give medical or legal advice, never quote a price you are not certain of, never argue with somebody who is annoyed.
Give it the sentence to say when it does not know
Word for word. "I am not certain about that — let me get someone who is."
This is the highest-value line in the whole field. An agent with no instruction for not knowing will fill the gap confidently, and confident wrong answers are the single most damaging thing an AI agent does.
Constrain the length
"Keep replies under 3 sentences." On chat this stops walls of text; on voice it is essential, because a long spoken answer gets interrupted and the caller loses the thread.
Set tone with an example, not an adjective
"Friendly but brief" means nothing to an agent. One sample exchange written the way you want it to sound does more work than a paragraph of description.
Keep the facts out of it
Prices, hours, policies and product details belong in a knowledge base, where they can be corrected without touching the instruction and can serve every other bot you build.
An instruction reciting prices from memory is the exact shape that produces confidently wrong numbers six months later.
Check it worked
Read the instruction aloud. If it sounds like something you could say to a new starter and have them do the job, it is right. If it reads like a specification with clauses, it is too long and the agent will follow parts of it inconsistently.
Then test: ask the five real questions, ask something it must refuse, and ask something it cannot know.
The parts people get wrong
- Changing the instruction changes the bot everywhere it is used. Read the Used-by list first.
- An instruction that tries to cover every possible question becomes a decision tree in prose and performs worse than a short one with a good knowledge base behind it.
- On a voice bot, read it aloud specifically. Text that scans well often speaks badly.