ChatropicA good support workflow needs more than a good first reply. It needs to ask for missing details, use the right tools, handle failures, and know when to bring in your team.
The new Skills experience brings that work together. Write a procedure in a document-style editor, ask Copilot to refine it, and use simulations to see how it behaves across a conversation. Here is how to get started, using a critical-issue escalation as our example.
One workflow, three connected tools
- Skills define the behavior. Give a customer need a repeatable procedure, with instructions, conditions, tool calls, and handoffs.
- Copilot helps author the procedure. Describe a change, inspect the proposed instructions, and decide whether to apply them.
- Simulations put the procedure to work. Define a customer situation, supply sample tool responses, and check the resulting conversation against your expectations.
Start with one request your team understands well. For our example, the goal is simple: gather enough context about a critical issue, notify the support team through an available Slack tool, and confirm the notification only when it succeeds.
1. Write your first skill
Open your agent in Agent Builder, select the specialist responsible for the request, and open or add a skill. Give it a name that describes the outcome, such as Escalate critical issue to Slack.
- Fill in When to use this skill. Describe both the customer need and the cases it should exclude.
- Write the steps in the document editor. Press Enter to add a step, or use the insertion controls for a condition or another step type.
- Type @ to insert an available tool. Check that the specialist has access to the operation you need.
- Add the behavior for missing information and unsuccessful actions. State what the agent can truthfully tell the customer at each stage.
A customer reports a critical issue that the available support tools cannot resolve. Exclude routine questions and non-urgent requests.
A useful first draft has four steps: identify the issue, ask for missing details, send the escalation through your connected tool, and confirm only after success. Add an explicit failure branch so a tool error cannot turn into a false promise.
Use Knowledge to check the policy behind the procedure. A tool name does not tell the agent your escalation policy, and writing a tool into a skill does not connect that service.
2. Improve the draft with Copilot
Select Copilot in the skill editor. The assistant can work with the current procedure, help clarify its instructions, and suggest simulation cases. Give it a specific change and the behavior you want to preserve.
Review this escalation skill. Ask for the affected service and a short issue description before notifying support. Keep tool errors separate from successful results. If the Slack tool fails, explain that the notification could not be confirmed and offer our approved fallback. Never invent a response time.
- Read the response and open Review proposed procedure when Copilot returns a skill change.
- Check the tool names, conditions, policy details, and fallback instructions.
- Choose Apply changes to update the draft. You can keep refining it before saving.
For test coverage, ask: “Create simulation cases for a successful escalation, missing details, and a failed notification.” Choose Review simulation setup on a proposed case to inspect the criteria and sample data before saving or running it.
3. Build a realistic simulation
Select Test skill → New simulation. You can also begin with a Happy path, Missing information, or Ambiguous request starter. Give each scenario a name that tells you what it checks.
- Write the customer’s opening message. Use the words a customer would actually send, such as “Something is seriously wrong. I need urgent help.”
- Add details the customer can reveal. In Additional details from customer, explain what they know and how they should respond when asked. These details are initially known only to the simulated customer.
- Set the context. Choose a simulated customer and channel. Add customer attributes if the agent should know them at the start.
- Configure tool responses. For available tools, choose sample success, empty, or error outcomes. Use response fields that match your connected tool. The agent sees that sample when it calls the tool.
- Define success. Use reply, tool, or outcome criteria as appropriate, and describe the overall successful conversation.
- Check Run settings, then Save or Run. Allow enough turns for the customer to answer a clarification question and for the agent to finish the procedure.
Keep the distinction between customer knowledge and agent knowledge clear. If you put every missing detail in the initial customer attributes, you will not test whether the agent asks the right follow-up question.
Simulation tools return your configured sample data; connected tools are not executed. Use a separate integration check to verify real credentials, permissions, and the live service response.
Four scenarios to try
A critical issue with enough detail
Customer: “Checkout is failing for every customer. It started ten minutes ago.” Configure a successful notification result. Expect the agent to gather any remaining required context, use the intended tool, and confirm the notification after success.
An urgent message with missing information
Customer: “Something is seriously wrong. Help!” Put the checkout failure in Additional details from customer, with instructions to reveal it only when asked. Expect a clarification question before escalation. Review the conversation order as well as whether a tool was called.
The notification service fails
Use the same critical issue, but configure an error response for the Slack tool. Expect a truthful explanation and your approved fallback. The agent should not tell the customer that support has been notified.
A routine request that should not escalate
Customer: “What time does support open tomorrow?” Supply the relevant approved knowledge. Expect a normal answer or clarification, with no critical-issue notification. A tool criterion can check that the escalation tool was not called.
Replace escalation with an order lookup. Test a valid order, a missing order number, an empty lookup result, and a service error. An empty result means no matching order was returned; an error means the lookup did not complete. Give each a different response.
4. Review, improve, and replay
Run a scenario individually or use Run all for a suite of up to ten scenarios. Open a result and read the Conversation, the criteria evidence, and Tool activity. Check what the customer said, what the agent knew at that moment, and whether the tool activity supports its reply.
Reply checks are AI evaluated; tool-call checks use recorded activity. A passing label is useful evidence, but it is still worth reading the exchange, especially when tone, timing, or an exception matters.
On a failed result, choose Improve with copilot to bring the test evidence into a focused revision. Review the proposed change, apply it to the draft, and select Replay scenario. Re-run the other cases too, so fixing a failure does not break the successful path.
If a result is marked stale after the skill, scenario, persona, or configuration changes, run it again. The old result describes the earlier setup.
5. Save with confidence
Applying a Copilot proposal changes the draft. Running a simulation does not publish the skill. Save is a separate action, and saving changes can update the skill used by enabled channels. Review the skill’s Active setting and the current environment before saving.
Before customers use the workflow, check the specialist’s knowledge and tool access, confirm the real integration works, and verify the missing-information and failure cases. Keep those simulations: they become a repeatable check whenever your policy or procedure changes.
Start with one skill and a handful of scenarios. The aim is a procedure your team can read, explain, and improve using the evidence from each test.
New to the workspace? Read the Agent Builder guide for the wider setup flow.