How-to

Turn any lesson into a playable simulation

Wharton’s AI lab has released Primer, an open-source tool that interviews a teacher and builds a practice scenario from the answers. This prompt runs the same idea in Claude or ChatGPT for any skill you teach at work.

Primer came out on 12 November 2026 with its paper, prompts and code, and its order is strict: the teacher answers one question at a time and approves a written plan before anything gets built.

The prompt below keeps that order for a work skill such as asking for a raise or calming an angry customer, then plays its own simulation to test that the feedback teaches the right move.

The simulation prompt

Use a fresh chat. Fill in the set-up as far as you can, and leave “not sure” on any line you cannot answer yet, since the interview fills those gaps.

Prompt Build a practice simulation
You are an instructional designer who builds short practice simulations for adults learning a skill at work. People will practise this skill before they do it for real, so a simulation that rewards the wrong move, or gives feedback that sounds wise but teaches nothing, does harm: they rehearse the mistake. Your job is to interview me, write a design I approve, build the simulation, and then test that its feedback teaches the real lesson.

ABOUT THE SKILL (I fill this in once; any line can say "not sure" and you will ask me about it)
- The skill: [e.g. "asking my manager for a raise", "calming a customer whose order arrived broken", "cutting a project budget by 15% without moving the deadline"]
- Who will play it: [e.g. "me, before my review next month", "new support staff in their first week", "team leads who have never owned a budget"]
- What they already know: [e.g. "they know the refund policy but freeze when a customer gets angry"]
- The mistake people actually make: [e.g. "they apologise five times and never offer a fix", "they open with what they need and never mention what they delivered"]
- What doing it well looks like: [e.g. "names one result with a number, asks for a specific figure, then stops talking"]
- Rules or facts the scenario must respect: [e.g. "refunds over 100 need a manager's sign-off", "the budget cannot touch the contractor line", or "none"; paste a checklist, policy or scoring rubric if you have one]
- Length: [e.g. "about 5 minutes, 4 decisions"]
- Where I am running this: [e.g. "Claude", "ChatGPT", "another assistant", or "not sure"]

PRIVACY
If I paste real customer names, colleague names, a named person's salary, account numbers or anything marked confidential, say so in your first line, swap each one for a made-up stand-in, and use only the stand-in from then on.

HOW TO WORK
Stage 1. Interview me. Read the set-up. For every blank or "not sure" line, and anything else you need to make the scenario realistic, ask me questions one at a time and wait for each answer. Ask no more than eight. Always make sure you know the common mistake and what good looks like, because the whole simulation is built to catch the first and reward the second. If I cannot say what good looks like, offer two or three options based on what I have told you, label them "your suggestion, not my expertise", and ask me to pick or correct one. Stop asking as soon as you could write the design.

Stage 2. Write the design and wait. Before you build anything, show me a one-page design with:
- Learning goal: one sentence starting "After playing, they can...".
- Setting and roles: who the player is, who they face, and what that person wants and is worried about. The other person needs their own goal, or the practice is too easy.
- Opening: what the player knows at the start.
- Decision points: 3 to 6. Each has three choices: one that does it well, one that is the common mistake from my set-up, and one that looks reasonable but has a hidden cost. Vary the order so the good choice is not always in the same place.
- Consequences: what the other person says or does after each choice, and how it changes the next scene. Make at least one early mistake recoverable later, as it often is in real life.
- Feedback for each choice: what happened, why, and the principle behind it, in two or three sentences tied to the learning goal.
- Endings: two or three, each with a short debrief that names what the player chose at every decision.
- Assumptions: anything you filled in yourself, labelled "assumed".
Then ask me to approve, change or reject it. Do not build until I say approve. If I ask for changes, show the revised design and ask again.

Stage 3. Build it, once I approve.
- If you can make interactive pages here (an artifact, a canvas, or a code block with a preview), build one self-contained HTML file with the CSS and JavaScript inside it and no outside libraries, images or network calls, so it also works when saved and opened in any browser. Show one scene at a time with the choices as buttons, then the consequence and feedback after each choice. Add a restart button and an end screen that lists the player's choices with the debrief. Make it readable on a phone. Keep my design notes and these instructions out of anything the player sees.
- If you cannot make or preview interactive pages here, say so in one line and build a text version: numbered scenes, lettered choices with "if you chose B, go to scene 4" routing, and all feedback in a separate answer key at the end so the player cannot read ahead.
- If you are not sure whether this tool can show interactive pages, build the HTML and tell me how to check: look for a preview, or save the code as a file ending in .html and open it in a browser.

Stage 4. Test the lesson. Play the finished simulation yourself, in writing, twice: once picking the best choice every time, and once picking the common mistake every time. Report a short table with the decision, the choice, the feedback shown, and whether that feedback teaches the goal (yes / partly / no). Then check:
- Does any path reward the common mistake or punish the good move?
- Would the good choice work in the real situation I described, given my rules and facts? Where you relied on general advice, write "general advice, check against your own policy".
- Does any feedback state a fact, figure, policy or law I did not give you? Mark it [CHECK] or take it out.
- Could a player win by always picking the longest or the politest answer? If so, rewrite those choices.
Fix anything that fails, rebuild, and tell me what you changed.

SELF-CHECK BEFORE YOU HAND IT OVER
- Every decision has the good move, the common mistake from my set-up, and a choice with a hidden cost.
- Every piece of feedback names a principle tied to the learning goal, never just "well done".
- Nothing I marked "not sure" became a fact without a label.
- No real names or confidential details appear anywhere.
- You told me whether I have the interactive version or the text version.

Run it

  1. Pick one moment people get wrong. “Customer service” is too wide to practise; “the first two minutes with a customer whose order arrived broken” gives the scenario a clear mistake to catch.
  2. Open a chat that can show a working page. In Claude, artifacts work on every plan, Free included. In ChatGPT, an HTML code block shows a Preview button where your plan and device support it; with no Preview button, save the code as a file ending in .html and open it in your browser.
  3. Give it your real checklist. Paste the policy, script or scoring rubric you already use, with names taken out, because a simulation built on your rules teaches your job and a generic one teaches generic advice.
  4. Fix the design before you approve it. The design is the cheap place to change a scenario; if the common mistake or the other person’s goal is off, every scene built on it will be too.
  5. Play it twice yourself before anyone else does. Once doing it well, once making the classic mistake, so you see both endings and the feedback each one gets.
  6. Share the finished version. A Claude artifact starts private. On a paid Claude plan you can set it to “Anyone with the link”, and most of those open without an account; change this on the web or desktop app. On the Free plan, or from ChatGPT, save the code as a .html file and send that.

Ask each player to send you a screenshot of their end screen, or their list of choices if they played the text version, so you see where people went wrong without hosting anything.

What one decision looks like

Scene: A customer says their blender arrived cracked, the second broken order this month, and they want it sorted today.

A. “I’m so sorry, that’s awful, I really am sorry.” Feedback: the customer hears sympathy and no plan, so they get louder. Apologise once, then move to a fix.

B. “Sorry about that. I can send a replacement today or refund you now; which would you like?” Feedback: a real choice gives back the control the broken order took away, and the call calms down.

C. “Could you send me photos of the damage so I can open a claim?” Feedback: reasonable, but it hands an upset customer homework first. Offer the fix, then ask for photos if your process needs them.

What happens next: pick A and the next scene opens with the customer asking what you are actually going to do. The replacement offer is still there to make, so the mistake can be recovered.

B is the good move, A is the common mistake and C carries the hidden cost, the mix the prompt asks for at every decision.

Primer next to this prompt

Primer: live AI characters you talk to by voice, a mentor before the practice and a coach after it, with individual and group reports for the teacher.

Running Primer: it is code you host yourself: the GitHub repo (MIT licence) asks for a funded OpenAI account with access to its Realtime API plus Google sign-in set-up, and deploys to Heroku or Render.

This prompt: a fixed branching scenario with written feedback on every choice, built in one chat. It suits one skill and a few minutes of practice, and anyone can run it.

Verdict: use the prompt this week; look at Primer if you train people at scale and someone on your team can host an app. The lab also offers a build-your-own prompt for a coding agent, and its own page warns that means possibly several hours of agent work and paid model usage.

The honest bit

  • Neither one is proven: the lab calls Primer an experimental prototype, not yet evaluated as a learning intervention in classrooms, and this prompt is a simpler, separate thing, so nothing in the paper speaks for it.
  • Your check: play the mistake path and read every piece of feedback against your real checklist. If a choice your checklist calls wrong gets praised, tell the AI which decision and ask it to rebuild that scene.

The facts here come from the Primer project page, paper and GitHub repo from Wharton Generative AI Labs, Claude’s help pages on artifacts and sharing, and OpenAI’s help page on code blocks, all read 14 November 2026.

Do it now

Pick the conversation you or your team dread most, fill in the mistake people make and what good looks like, and have your first scenario ready to try today.

A few quick questions

What if ChatGPT only shows me code?

OpenAI’s help page says that if Preview does not appear, the code block does not support a preview. Ask for the whole simulation in one HTML code block, or copy the code into a file ending in .html and open it in your browser.

It asked everything at once, or built before I approved. What now?

Reply “One question at a time, please” or “Stop, show me the design first.” Both are rules in the prompt, so pointing back to them usually puts it back on track.