New Topic Agent Skill

Create New Topic

Generate a new Copilot Studio topic YAML file based on user requirements.

Instructions

• Auto-discover the agent directory:

   Glob: **/agent.mcs.yml

If multiple agents found, ask which one. NEVER hardcode an agent name.

• Check for matching templates in ${CLAUDE_SKILL_DIR}/../../templates/topics/ first:
• greeting.topic.mcs.yml — OnConversationStart greeting
• fallback.topic.mcs.yml — OnUnknownIntent fallback with escalation
• arithmeticsum.topic.mcs.yml — Topic with inputs/outputs and computation
• question-topic.topic.mcs.yml — Question with branching logic
• search-topic.topic.mcs.yml — Generative answers from knowledge
• auth-topic.topic.mcs.yml — Authentication flow
• error-handler.topic.mcs.yml — Error handling
• disambiguation.topic.mcs.yml — Multiple topics matched

If a template matches, use it as the starting point.

• MANDATORY: Verify ALL kind: values against the schema before writing them:

   node ${CLAUDE_SKILL_DIR}/../../scripts/schema-lookup.bundle.js kinds                    # List all valid kind values
   node ${CLAUDE_SKILL_DIR}/../../scripts/schema-lookup.bundle.js resolve AdaptiveDialog   # Resolve trigger structure
   node ${CLAUDE_SKILL_DIR}/../../scripts/schema-lookup.bundle.js resolve <TriggerType>    # Resolve specific trigger
   node ${CLAUDE_SKILL_DIR}/../../scripts/schema-lookup.bundle.js search <ActionKind>      # Verify an action kind exists

NEVER write a kind: value you haven't verified exists in the schema. This is the #1 source of hallucination errors. If schema-lookup.bundle.js search <kind> returns no results, the kind does NOT exist — do not use it.

• Determine the trigger type from the user's description:
• OnRecognizedIntent — For topics triggered by user phrases (most common)
• OnConversationStart — For welcome/greeting topics
• OnUnknownIntent — For fallback topics
• OnEscalate — For escalation to human agent
• OnError — For error handling

• Generate the topic YAML with:
• A # Name: comment at the top
• kind: AdaptiveDialog
• Appropriate beginDialog with correct trigger
• Unique IDs for ALL nodes (format: <nodeType>_<6-8 random alphanumeric>)
• If using a template, replace ALL _REPLACE placeholders with unique IDs

• Check settings.mcs.yml for GenerativeActionsEnabled. Read the agent's settings.mcs.yml to check.

• Save to the agent's topics/<topic-name>.topic.mcs.yml directory

• MANDATORY: Validate the generated file after saving:

Step A: Schema validation — always run this first:

   node ${CLAUDE_SKILL_DIR}/../../scripts/schema-lookup.bundle.js validate <saved-file.yml>

Step B: LSP-based validation — also run this if the agent has .mcs/conn.json:

Read .mcs/conn.json to get connection details, then:

   node ${CLAUDE_SKILL_DIR}/../../scripts/manage-agent.bundle.js validate \
     --workspace "<path-to-agent-folder>" \
     --tenant-id "<tenantId>" \
     --environment-id "<envId>" \
     --environment-url "<envUrl>" \
     --agent-mgmt-url "<mgmtUrl>"

Both validations are complementary: schema validation checks structural correctness (action kinds, property placement), while LSP validation checks Power Fx expressions, cross-file references, and environment-specific rules. If either fails, fix the issues before reporting success to the user.

Generative Orchestration Guidelines

When the agent has GenerativeActionsEnabled: true in settings:

Use Topic Inputs (AutomaticTaskInput) instead of Question nodes to auto-collect user info.

Place inputs at the AdaptiveDialog root level (NOT inside beginDialog):

kind: AdaptiveDialog
inputs:                    # <-- at AdaptiveDialog root, NOT inside beginDialog
  - kind: AutomaticTaskInput
    propertyName: userName
    description: "The user's name"
    entity: StringPrebuiltEntity
    shouldPromptUser: true
beginDialog:
  kind: OnRecognizedIntent
  id: main
  actions:
    - ...

• The orchestrator auto-collects inputs based on the description — no explicit Question node needed.
• Still use Question nodes when: conditional asks (ask X only if Y), or end-of-flow confirmations.

Use Topic Outputs instead of SendActivity for final results.

Use outputType at the root level and SetVariable to set output values — do NOT use TaskOutput (which is only valid in TaskDialog connector actions):

kind: AdaptiveDialog
inputs:
  - ...
beginDialog:
  kind: OnRecognizedIntent
  id: main
  actions:
    - kind: SetVariable
      id: setVar_abc123
      variable: Topic.result
      value: ="computed value"
outputType:                # <-- at AdaptiveDialog root
  properties:
    result:
      displayName: result
      description: The computed result
      type: String

• The orchestrator generates the user-facing message from outputs.
• Do NOT use SendActivity to show final outputs (rare exceptions: precise mid-flow messages).

Include inputType/outputType schemas at the AdaptiveDialog root level when using inputs/outputs:

inputType:
  properties:
    userName:
      displayName: userName
      description: "The user's name"
      type: String
outputType:
  properties:
    result:
      displayName: result
      type: String

Topic-Action Chaining Under Generative Orchestration

When a topic exists alongside other topics or other actions (i.e. TaskDialog), think carefully about the topic outputs — it directly affects whether the orchestrator will chain to an action.

Two scenarios:

• Topic is self-contained (e.g., reservation confirmation, calculation result): The topic does the work itself. It can gather the inputs too or just ask for those, but work is executed into the topic. In such case, output a confirmation/status message. The orchestrator treats the task as done or not, depending on the confirmation — this is correct.

• Topic gathers data for an action to consume (e.g., collecting a complaint that a Teams action should send, without manually calling the action into the topic): The topic output must be the data itself, not a confirmation. If you output "Your complaint has been sent!", the orchestrator assumes the job is done and will NOT invoke the action which means the complaint will never be sent. Instead, output the complaint text so the orchestrator can see there's an action to send it and feed it to such action.

Ask yourself: Does this topic complete the task on its own, or does it prepare data for an action/other topic? If the latter, output the data, not a status message.

Alternative approaches for data-gathering topics:

• Remove the topic entirely and tailor the action's input descriptions for the specific use case (simplest). Basically the action will gather the data directly via its inputs, no topic needed. Useful for simple data collection (i.e. one-shot).
• Use a global variable: the topic sets it (one-shot or by asking multiple questions and consolidating into the same variable via PowerFX), and the action's body references it via Power FX. This is more complex but can be useful for multi-turn data gathering that feeds somewhere else, especially if you need to do some logic on the data before sending it.
• Output the collected data in a topic output variable so the orchestrator can pass it to the next action/topic.

Power Fx Quick Reference

• Expressions start with =: value: =Text(Topic.num1 + Topic.num2)
• String interpolation uses {}: activity: "Hello {Topic.UserName}"
• Common functions: Text(), Now(), IsBlank(), !IsBlank(), DateTimeFormat.UTC
• Variable init: variable: init:Topic.MyVar (first assignment uses init:)
• Type coercion to text: Use SetTextVariable instead of SetVariable to convert non-text types (Number, DateTime, etc.) to text via template interpolation: value: "Guests: {Topic.NumberOfGuests}"