Example prompts

Once the Zowie connector is set up, you do not call tools by name. You say what you want, and your assistant works out which tools to use. The prompts below are written the way you would actually type them. Copy one, replace the bits in ANGLE BRACKETS with your own ids, dates and wording, and send it.

Ten groups, one per thing the connector is good at, with a few variants behind the tabs in each. Read the tab labels alone and you have seen the whole surface. What you can do is the detailed map.

Two habits worth picking up.

Say which environment and AI agent you mean, or tell the assistant to ask you when it is not sure. If you work with one environment only, it will usually find it by itself.

Ask for the reasoning, not just the answer. Adding "tell me which log events you based that on" turns a plausible summary into something you can check. For anything that changes your configuration, "show me the change before you write it" is worth adding too. Those edits land on the STAGING draft by default, which you review and publish yourself. Contact reasons are the exception: they have no draft version, so changing one is live straight away. See Permissions and versions.

One conversation: what happened and why

Here is an interaction ID: <INTERACTION_ID>. Tell me what happened in this conversation and why the customer ended up unhappy. Tell me whether this was a knowledge problem, a routing problem or an integration problem, and quote the evidence.

Interaction <INTERACTION_ID> called our order-status API. Show me what we sent, what came back, and how the agent used the response.

For interaction <INTERACTION_ID>, show me the exact prompt the model received right before it gave the answer the customer complained about, and tell me which part of our configuration put that text there.

Many conversations: finding the pattern

Runs on your ordinary access — none of this needs analytics switched on.

Analyse the conversations from the last 7 days that ended in a handover. Group them by what the customer was actually trying to do, and tell me the three biggest causes.

Analyse every chat interaction from yesterday that got a thumbs-down. Read the worst five and tell me, for each one, whether it was a misrouted intent, an API error, or a bad answer.

Something changed after we published on <DATE>. Compare interactions from the week before and the week after and tell me what got worse.

Improving a process

Improve the process <PROCESS NAME>: when the shipping API times out, the customer currently sees a raw error. Change it so they get a calm message saying we are checking with the warehouse, and hand over to a human. Leave everything else as it is.

Show me the <PROCESS NAME> process and explain it to me step by step in plain language, as if I had never seen a flow diagram.

Review the <PROCESS NAME> process for dead ends: places where the conversation can stop without an answer and without a handover. List them and propose a fix for each.

Tone and rules

Add a rule: when a customer mentions a chargeback or a payment dispute, the agent must not promise any outcome and must offer to hand over to a human.

Our agent sounds too casual on email. Show me the current persona, then make the email tone formal and concise while leaving chat as it is.

List every rule we have that mentions refunds. Are any of them contradicting each other?

Routing: what customers mean

Customers asking to change their delivery address keep landing in the returns flow. Find out why and fix the routing.

We keep getting questions about payment plans and we have no intent for it. Create one, route it to <PROCESS NAME>, give it the phrasings our customers really use, and leave it switched off so I can review first.

What the agent is allowed to say

Knowledge is read-only over MCP. Editing it still happens in the Zowie UI.

A customer was told we refund shipping costs. Is that in our knowledge base, and if so, where exactly?

List the knowledge articles about payment methods and tell me which ones look out of date or contradict each other.

Numbers and reporting

The only group that needs analytics switched on for your account — see Permissions and versions. The data is loaded once a day, so today is never in it.

Compare automation rate and average CSAT for chat in <MONTH> versus the month before, broken down by contact reason.

What are our top ten contact reasons by volume in the last 90 days, and which of them has the worst resolution rate?

Which contact reason grew the fastest in the last 60 days? Then read a few of those conversations and tell me whether we are handling them well.

Testing before you publish

A real conversation with the real agent on your draft, held inside the chat you are already in. It calls your systems for real, so treat it as a rehearsal.

Start a test chat on staging and act as a customer whose parcel is marked delivered but never arrived. Push back once when the agent asks you to wait, then ask for a refund. Show me the whole exchange and flag anything that contradicts our delivery policy.

Test the change we just made to the returns process: play a customer with an order from 45 days ago who wants to return a sale item. Tell me whether the agent handles it the way we intended.

Run five short test conversations on staging, each phrasing "I want my money back" differently, and tell me whether they all reach the same process.

Reporting categories

Contact reasons are the one thing with no draft copy, so a change takes effect at once. Conversations already classified keep the category they have.

Show me our contact reasons. Are any of them overlapping or too vague to be useful for reporting?

"Payment issue" and "Card declined" overlap. Merge them into one category with a description that clearly excludes refunds, and add a new "Subscription cancellation" category. Show me the full set before you write anything, and remind me this goes live immediately.

Telling Zowie something is wrong

Report this to Zowie: in the conversation we just looked at, the event log came back out of order and the transcript was missing the last agent message. Include everything we did in this session.

Report to Zowie that I cannot find a way to see which version of the configuration served a given conversation, and explain why I needed it.