Conversation design for LLM assistants: beyond the prompt
When the model can already write fluent replies, conversation design is deciding how the assistant behaves across a whole exchange: how it opens, when it asks and when it assumes, how it admits it does not know, how it recovers from a misunderstanding and how it hands over. The system prompt is where you write those decisions down. Transcripts are how you find out whether they hold.
What is left when the model is fluent
Conversation design used to mean writing every line a bot could say and drawing the flow chart that joined them. Language models removed that job. They did not remove the decisions behind it. Someone still has to decide what a good conversation is for this product and these users, and then check that it happens.
A useful yardstick is older than any of this. The philosopher Paul Grice described what cooperative speakers do in four maxims: say what is true, say as much as is needed and no more, be relevant, be clear. Most assistant failures break one of them. The made-up answer breaks the first. The four-paragraph reply to a yes-or-no question breaks the second. The helpful tangent breaks the third, and the wall of bullet points the fourth.
Our running example is the support assistant for a home broadband provider. Its users are often cut off from the internet, on a phone, and annoyed before they type a word.
Persona and tone rules that hold up
The weakest persona is a list of adjectives. “Friendly and professional” tells the model nothing it would not do anyway, and “empathetic” produces an apology at the top of every message.
Rules that hold up have three properties. They describe behaviour you could spot in a transcript. They give the reason, so the model can apply the rule to cases you did not foresee. And they were written with the bad day in mind: the angry customer, the third repeat of the same question, the request the assistant must refuse. Anthropic’s prompting guidance says much the same: explain why an instruction matters, tell the model what to do instead of what not to do, and drop the capital letters, since current models over-apply shouted rules.
Here is one section of a system prompt, before and after.
Before:
You are Aria, a friendly and professional assistant for Example Broadband.
ALWAYS be helpful. NEVER make things up.
Keep answers short. If you cannot help, escalate.
After:
You are the support assistant for Example Broadband. Customers reach you in
the app, usually on a phone and often while their connection is down, so
they are short of patience and may be on mobile data.
How to reply
- Lead with the answer or the next step. Add at most two sentences of context.
- Once the customer has named something ("the box", "the hub"), use their word.
- Apologise once, when the fault is ours, then move to the fix. Repeated
apologies read as stalling.
- If the customer is angry, drop the pleasantries and say what you can do now.
What you know
- Answer account and fault questions only from tool results and the help
articles supplied. If they do not cover it, say so and offer a person.
A wrong answer about a bill costs more than a handover.
- You are an AI assistant. Say so in your first message and whenever asked.
Asking and assuming
- If a wrong guess is easy to undo, go ahead and state what you assumed.
- If it would change a bill, an appointment or a contract, ask first:
one question, with the likely options.
Handing over
- Offer a person when the customer asks, after two failed attempts, and
for complaints, bereavement or anything legal. Say who will reply and
when. Only say you are transferring if someone is there to take it.
The second version is longer, and every line in it can be checked against a transcript. Add three to five example replies, including one to an angry customer and one refusal, and the tone settles further.
How should it open, ask and assume?
Opening messages
The opening sets expectations for everything after it. “Hello, I’m your virtual assistant, how can I help you today?” wastes it.
A good opening does up to four things in two or three sentences: says it is an AI, says what it can do, uses what is already known, and invites the customer to write in their own words. For the broadband assistant during a known fault: “I’m Example Broadband’s AI assistant. There’s an outage in your area that engineers expect to fix by 3 pm. Is that why you’re getting in touch, or is it something else?” The best opening line is often an answer.
Do not open by asking for an account number before you know whether the question needs one.
Clarifying questions or assumptions
Each clarifying question costs the user a turn and gives them a reason to leave. Each wrong assumption costs a correction, or worse. So the rule turns on what a wrong guess costs and whether it can be undone. OpenAI’s Model Spec puts the trade-off in one line: “The expected benefit of clarification should justify its cost.”
In practice:
- Cheap and reversible: assume, and say so. “I’ve assumed you mean the main line on the account. Tell me if it’s the other one.”
- Costly or permanent: ask one question and offer the likely answers. “Do you want to move the engineer visit, or cancel it?”
- Never ask for something a tool can look up, or a question whose answer would not change what you do next.
When several details are needed, let the customer give them in any order and confirm once at the end, instead of running a form one field at a time.
Not knowing, repair and escalation
Saying “I don’t know”
A usable “I don’t know” has three parts: what the assistant cannot answer, why, and what happens next. “I can’t see why that charge was added, because the billing notes I have don’t cover it. I can pass this to the billing team, who reply within one working day. Shall I?” Without the third part it is a dead end.
Keep three situations distinct in the prompt, because the right wording differs: I do not have that information, I am not permitted to do that, and I did not understand you.
Repair after a misunderstanding
The signals of a misunderstanding are easy to recognise: the customer rephrases, repeats, writes “no”, or gets shorter and sharper. The assistant should own the miss briefly, restate what it now understands, and try a different approach instead of giving the same answer again.
Customer: no, I said the NEW address
Weak: I apologise for the confusion. To update your address, go to
Settings, then Account.
Better: My mistake. You've already moved, and the bill is still going
to the old address. I can correct the billing address now.
Is it 14 Mill Lane?
On the second failed repair, change strategy: offer options to tap, or a person. And the fault is never the customer’s. “I didn’t understand your question” becomes “I got that wrong”.
Escalation language
The trigger for a handover is half the design. The wording is the other half. “I’m unable to help with that, please contact support” is absurd when the customer is already in support.
A good handover message says what will happen, who will do it, when, what has been passed on and what the customer can do meanwhile. Give honest timings: “a few minutes” that turns into four hours is worse than “by tomorrow morning”. Never make someone ask twice for a person. The triggers, routing and summaries are covered in our guide to AI customer support.
Memory across sessions
Memory helps when it saves the customer from repeating themselves and harms when it surprises them. What we store: stable facts and preferences (language, preferred channel, router model), open issues, and promises made (“we said we would call back on Friday”). What we do not: sensitive details offered in passing, and anything about mood.
Three rules. Date everything, and check anything old before relying on it (“Are you still at Mill Lane?”). Let users see and delete what is remembered. And keep memory in your own systems, per user. Anthropic’s memory tool, for example, has the model request reads and writes while your application does the storing, and its documentation recommends stripping sensitive data and expiring stale entries.
Voice versus text
| Text | Voice | |
|---|---|---|
| Length | Short paragraphs the user can scan | One or two sentences a turn |
| Choices | Buttons or a short list | Two or three options at most, spoken |
| Details | The user can scroll back | Read back names, dates and numbers |
| Timing | The user sends when ready | Pauses and interruptions are part of the talk |
| Formatting | Links and lists work | None; send links by message |
| Waiting | Typing indicator | A spoken “let me check”, then the result |
Voice prompts therefore need their own rules. OpenAI’s prompting guide for its live voice model includes two we would put in any spoken assistant: keep listening while the caller pauses to think, and do not state a price or confirm a booking until the backend has returned it. The engineering behind spoken assistants is in our guide to voice AI agents.
Testing with transcripts
You cannot review conversation design by reading the prompt. Read conversations. Each week take a sample that deliberately includes handovers, conversations the customer abandoned, long ones and a random slice, and mark each against the same list:
- Did the opening say what the assistant is and what it can do?
- Did the first reply answer the question or move it forward?
- Was every factual claim backed by a source or a tool result?
- Were the questions necessary, and asked one at a time?
- Where it assumed, did it say so?
- When it did not know, did it say so and give a next step?
- After a misunderstanding, did it change approach?
- Was the length right for the channel?
- Did the tone follow the customer’s state?
- Did the handover say who, when and what was passed on?
- Did it stay within scope and policy?
- Would you be content to have sent these replies yourself?
Every failure becomes a test conversation that runs before the next prompt or model change. Change one thing at a time, and version the prompt like code. That test suite is the start of proper evals and guardrails.
How to start
In the first week, pull 50 real transcripts. If there are none yet, have five colleagues from support play customers for an hour. Mark them with the checklist, find the two questions that fail most often, and rewrite those parts of the prompt as behaviour with reasons, plus a handful of examples. Then run the same conversations again.
The first decision is who owns the weekly review, because a prompt that nobody rereads against real conversations drifts out of date as the product changes. That review habit is a large part of the conversational AI work we do.