Realism is about cues, not decoration
Learners need to practice recognizing WHEN a skill applies, so the scenario should present the same cues and pressure as the real moment: the same kind of opening line, the same incomplete information, the same time pressure. Add detail that changes decisions; cut detail that's just set dressing. If real customers never announce "I am upset," the character shouldn't either.
Tuning difficulty
- Calibrate resistance to the audience: new hires get a firm-but-fair character; experienced staff get the hard case they actually struggle with.
- The character should reward genuinely good moves and hold its ground against weak ones. It should never collapse at the first reasonable sentence, and never be unwinnable.
- Use behavior triggers for deliberate complications: escalate if the learner stalls, introduce a curveball mid-conversation, soften only after a specific behavior appears. One or two triggers per simulation is plenty; more makes the character erratic.
- Difficulty should come from the situation (a realistic objection, a competing priority), not from the character being obtuse or withholding basics a real person would state.
One skill per sim; series for capabilities
A simulation practices one skill well. When the goal is a capability (run a full sales cycle, manage a performance conversation end-to-end), design a series:
- Sequence easy-to-hard: first sim isolates the core move with a cooperative character; later sims add resistance, ambiguity, and competing objectives.
- Vary the surface (different customers, products, moods) while keeping the skill constant. Varied practice is what makes the skill transfer.
- A coach between attempts (or a debrief after each) turns the series into deliberate practice rather than repetition.
Tie behavior to the learner, not the clock
Write triggers and instructions against what the learner does, never against turn, exchange, or message counts ("On turn 3, admit the billing error"). The character tends to treat turn numbers literally: it can trigger a beat before the learner has done anything relevant, or skip it entirely when the conversation runs shorter or longer than expected. The result feels scripted instead of responsive.
Tie the same beat to a learner action instead: "Once the learner asks what changed on the bill, admit the billing error." Now the beat is tied to the moment it belongs to, no matter how the conversation actually unfolds. This applies to the character description, character instructions, scenario context, and behavior triggers alike. The simulation editor shows an inline warning under any of those fields that contains turn-numbered language, so you can catch it before publishing. The warning is advisory and never blocks a save.
Choosing text vs voice
- Voice suits skills where tone, pacing, and thinking on your feet are the point: calls, de-escalation, interviews, pitches. Text suits skills where the learner benefits from composing and reconsidering: written support, negotiation over chat, or early-stage practice before voice pressure.
- Voice sessions use credits for every minute of real-time audio, on top of the AI generation credits, and every voice simulation has an enforced maximum session length that you set in the simulation editor. Size the cap to the scenario's natural length: a discovery call needs more room than a quick objection drill. The cap is also the cost lever for voice.
- A common pattern: learners rehearse in text first, then take the voice version of the same scenario as the assessment.