Teacher Describes a Lesson, AI Generates a Playable Simulation: Wharton Open-Sources Primer

Wharton Generative AI Labs has open-sourced Primer. Teachers do not need to write code; they go through a structured interview with AI to clearly explain course goals, student levels, common misconceptions, and assessment criteria. The system first provides a reviewable design document, and only after the teacher approves it does it generate a playable simulation.

Simulation-based teaching is effective, but expensive to produce. Designing a simulation that has students retrieve knowledge, explain the situation, act, see the consequences, and then revise used to take a great deal of time. Default LLM settings, moreover, help students too quickly and drift away from learning objectives; role-play may look plausible but lack instructional structure.

Primer's approach is to embed pedagogy into each agent's instructions. In the authoring phase, an Intake Agent interviews the teacher, a Design Agent writes the design document, a Casting Director generates character prompts, an Image Generation Agent produces scene images, and an Experience Builder assembles everything into an experience that can be assigned to a student or class. In the runtime phase, there are three student-facing agents: mentor, practice, and feedback. Practice is split into Roleplay and Teachback: the former has students navigate interactions with counterpart characters who have motivations and constraints, while the latter has students explain concepts to someone who does not yet understand them, then revise based on follow-up questions.

The most noteworthy part is the design-document approval step. Teachers first read the learning objectives, scenario flow, characters, stakes, and adaptation logic; they revise and approve it as if reviewing a lesson plan, and only then do downstream agents get to work. X user Raúl Nogales said that the reviewable design document is the most important thing: teachers can inspect the pedagogy before students see the simulation, so the simulation turns from a black box into a tool teachers can own.

Some commenters also poured cold water. X user 'Unknown' questioned whether what Primer generates is a projection of the teacher's existing understanding, and argued that the real difficulty is diagnosing students' incorrect mental models; opening the code matters more than the pedagogical promises. This is not a light criticism. It points out that there is still a gap between generating a simulation and generating effective feedback. Openly releasing prompts, templates, and code at least makes that gap inspectable, replaceable, comparable, and modifiable.

How to use it. The GitHub repository provides Heroku and Render deployment buttons and requires an OpenAI API key and Google OAuth configuration; a Claude key is optional. After logging in, install the Intake Agent, Roleplay Template, and Teachback Template on the Modules page, then create or play an experience. Developers can change the model, orchestration, and storage; the documentation covers LMS integration, PostgreSQL plus S3, OAuth/SAML, and configuration for Gemini or local Ollama.

The project explainer video is on YouTube. The project page is at gail.wharton.upenn.edu/primer, the paper is on SSRN, and the code is on GitHub. In the paper's appendix there is a Primer Creator Prompt: give the paper to an agent and let it build its own version of Primer; it may take several hours and incur paid model calls. This is probably the most unconventional part of the release: the paper is not only readable but executable.

The boundaries also need to be made clear. Primer is an experimental prototype; it has not been evaluated as a learning intervention under classroom conditions, nor has it undergone compliance review for FERPA, COPPA, state student privacy laws, or school district policies. Institutions that use it with students must take responsibility for their own review. Education products cannot rely on launch events.

The project's name comes from the Primer in Diamond Age. The Wharton team built the initial version together with teachers; the system remains neutral with respect to models and technology and can be modified as needed. GitHub uses the MIT License. For educators, it lowers the barrier to turning instructional judgment into simulations; for developers, it provides a modular agent orchestration; for researchers, it puts the prompts and the paper on the table together.