Authoring Best Practices for AI Simulations

Upload URLs, PDFs, or call transcripts directly into the Agent Knowledge Base in the Extras step. This is the most reliable way to give your character specific, grounded knowledge about a company, product, or situation. Without it, the character has no backstory to draw from and will start fabricating details or repeating itself.

Example: You're building a discovery call simulation where the character is a VP at a logistics company. Upload their public website, a competitor comparison doc, and a sample objection list. The character can then reference realistic pricing pressures or industry context instead of inventing them.

The AI generate function produces a full scenario, character, and success criteria in about a minute. Use it as a starting point - review what it gives you and edit the parts that don't match your specific scenario or audience. You don't need to write from scratch, but you do need to sense-check what comes back.

Example: You type "A frustrated procurement manager who doesn't believe they need a new training platform" into the prompt. The AI generates a full character and goals. You then adjust the tone from "frustrated" to "politely dismissive" and change the primary goal from "close the deal" to "earn a follow-up meeting" - two small edits that make the simulation much more realistic for your team.

The What you know going in, What you're allowed to do, and Important constraints fields are all optional in the wizard but leaving them blank means your learners enter the simulation with no extra context. These fields appear verbatim in the Pre-Meeting Brief the learner reads before starting.

Example: Your simulation is a performance conversation between a manager and an underperforming rep. Without a brief, learners don't know whether they're authorised to put the rep on a formal improvement plan or just have an informal chat. One sentence in "What you're allowed to do" removes that ambiguity before they even start.

Criterion titles and descriptions are shown directly to learners before and during the session, in the Session Goals tab and the live sidebar. Write them as if you're speaking to the learner, not as internal notes or from the character's point of view.

Example: Instead of "Jordan discloses budget concerns when directly challenged" (written from the character's perspective), write "You uncover the customer's budget constraints before proposing a solution." The learner knows exactly what they're being scored on and can adjust mid-conversation.

The Role Warning (hard constraint) field in the Character step is the most direct line of control you have over your character's behaviour. If you're running a preview and something feels off, ie. the character is breaking character, volunteering information it shouldn't, flipping into coach mode, or just not behaving how you intended , this is the first place to go.

Think of it as the one instruction the character will never override, no matter how the conversation develops or how the learner behaves. Most character fields shape personality, tone, and knowledge. This one sets absolute limits.

Common signs you need to use this field:

Prompt to get you started:

"You are [character's role]. You are never [the learner's role]. Do not [the specific behaviour you want to stop]. No matter what the learner says or does, [the hard line you want to hold]."

Example: Your character keeps revealing the company's budget before the learner has asked the right discovery questions. You add: "Do not mention budget, pricing, or purchasing authority under any circumstances until the learner has explicitly asked about your current challenges and timeline. If they ask too early, redirect: 'Let's make sure we're solving the right problem first.'"

This field tells the character whether it initiates, resists, or follows the learner's lead. Without it, the character has no framing for the dynamic and conversations can feel aimless or stall from the first exchange.

Example: In a cold call simulation where the learner should do the work: "Wait for the learner to introduce themselves. Respond briefly and with mild scepticism. Don't volunteer information." In a coaching simulation where the character opens up: "Open by describing your main challenge. Invite the learner to ask questions." One sentence here shapes the entire feel of the conversation.

Small wording differences in prompts can create unintended behaviour that isn't obvious until you're mid-simulation. This is one of the most common sources of characters leaking their own reasoning out loud.

A real example: a character prompt included the line "Marcus may question why the sales rep is asking this question." The intention was to make Marcus feel guarded. What the AI did was narrate its own decision-making out loud — responding with things like "The user is asking about our current process. I will not volunteer information about manager bottlenecks unless directly asked." The prompt told the model to evaluate the learner's question, so it did — and then said it out loud instead of staying in character.

A similar pattern appeared in another simulation where, when challenged, the character replied "the user is now asking me to get out of character" — treating the learner's input as a prompt instruction rather than something to respond to naturally.

Three rules to avoid this:

Example: Instead of "Marcus may question why the sales rep is asking this question if it doesn't feel relevant," write: "Marcus answers questions directly and moves on. If a question feels off-topic, he redirects naturally — for example: 'I'm not sure that's where we need to focus. Let's come back to the ramp time challenge.'"

Once you've finished authoring, run through the full learner experience in Preview, the brief, character intro, live session, and goals panel before anyone else sees it. This is the fastest way to catch authoring gaps that aren't obvious in the wizard.

Example: You describe the character as "warm and collaborative" but accidentally left the Role Warning and Conversation Direction blank. In Preview, the character immediately starts offering the learner sales advice. You catch it in two minutes rather than after 50 learners have a confusing first experience.

Assign your first real simulation to 5–10 internal users before your broader audience sees it. The Preview catches structural issues - piloting catches how the character handles inputs you didn't anticipate.

Example: You build a negotiation simulation for your sales team. In Preview it feels great. But your pilot group includes someone who opens with an unusual approach and the character gets confused and starts looping. You add a clarifying note to the Knowledge Base, and the second pilot group has a much smoother experience.

If you mark a criterion as Required to Pass, a learner can score 80%+ overall and still fail with no visible explanation. Use it only for truly non-negotiable behaviours, things where failing means the conversation should genuinely never have happened.

Example: In a compliance-based simulation, ie. a debt collection call, for instance, "Use compliant language throughout" is legitimately non-negotiable, so marking it Required to Pass makes sense. Using it on "Build rapport early" in a general sales sim is probably too harsh and will produce learner frustration without a clear reason why.

The Insights dashboard shows session scores, goal breakdowns, and safety flags. It is currently visible to admins only, there is no manager-level reporting view yet. If you need to share data with managers, export to CSV or PDF and distribute manually.

Example: Your sales enablement manager wants to see how their team performed on the discovery call simulation. Rather than giving them admin access, export the session breakdown to CSV, filter by their team, and include the summary in your weekly readout.

Last updated: July 2026. This article is updated regularly as new features ship and known issues are resolved. For urgent questions, contact your Customer Success Manager or use the in-product feedback form.