Design a role-play
A strong SCL role-play aligns the learning objective, the simulated customer’s behaviour, the learner briefing, the time rules, and the evaluation criteria. If any of those parts describes a different conversation, the experience may feel realistic but produce misleading feedback.
The current product has no authoring interface. Role-play records and content files are maintained through backend processes, and only active role-plays with an active SCAI agent are available to learners.
Start from the learning outcome
Define the change you expect to observe after practice:
- Who is the target learner?
- In which real situation will they use the skill?
- What should they do, say, decide, or avoid?
- What evidence in a transcript would demonstrate success?
- What should they learn from a weak attempt?
Keep one coherent conversation outcome per role-play. If the scenario requires unrelated skills, split it into multiple exercises or make the weighting explicit.
Author the learner briefing
| Content | Functional purpose | Authoring guidance |
|---|---|---|
| Name and scenario | Orient the learner | Explain the situation and relevant prior context without revealing the bot’s entire script |
| Objectives | Define success | Use observable outcomes rather than broad aspirations |
| Challenges | Prepare for difficulty | Describe emotions, objections, ambiguity, or constraints the learner must handle |
| Key tips | Support transfer | Give practical behaviours without prescribing a word-for-word answer |
| Skills | Label the capabilities | Use terminology that matches the evaluation indicators |
| Competency | Set expected difficulty | Choose basic, intermediate, or advanced consistently across content and scoring |
| Character profile | Make the counterpart credible | Define role, organization, language, and the context relevant to the conversation |
| Interaction | Set expectations | Describe interaction type and contact style in learner-friendly language |
| Help | Provide in-session support | Keep it short enough to consult without abandoning the conversation |
Each supported locale needs a complete, reviewed content entry. Avoid mixing languages inside one locale and do not rely on machine fallback for high-stakes learning material.
Set session rules
- Choose strict when ending at the planned duration is part of the skill or operational constraint.
- Choose flexible when a natural close matters more than the target duration.
- Set the warning early enough for the learner to react meaningfully.
- Set an explicit overtime limit for flexible sessions; otherwise the frontend falls back to five additional minutes.
- Decide separately whether elapsed time should contribute to the evaluation score.
Align the simulated customer
The SCAI character should have a stable identity, motivation, emotional state, knowledge boundary, and conversation path. Its difficulty and objections must make the briefing objectives observable without coaching the learner toward the answer.
The current operating process starts with a design-team handoff and a request to the SCAI team. A default customer-service agent is not suitable: the SCAI team creates a specialized Role Play agent, disables the default customer-service macro-prompt, enables a DNIS, and configures the approved ElevenLabs voice. The AI must remain the customer or counterpart while the human learner remains the employee being trained.
The role-play record links to one active SCAI agent. That agent supplies the voice environment and destination used by the browser and the AI Factory endpoint used by the evaluation backend.
Define the evaluation
Evaluation criteria are role-play-specific Markdown content stored separately from the learner-facing briefing. They define the phases or indicators, evidence rules, scoring rules, and weighting applied to the transcript.
Continue with Conversation and evaluation design for the assessment model, Create and provision a role-play for the current manual setup, and Publication and governance for release controls.
Design workflow
The approval and activation mechanism is organizational rather than product-managed today. Record owners, versions, evidence, and rollback information outside SCL until a governed catalog workflow is implemented.