Why We Built Primer
Students who can define a concept or recognize a correct answer often struggle when the same material appears in a new setting. Educational simulations offer the opportunity to bridge that gap. A good one asks students to retrieve what they know, interpret a situation, act, and revise in response to consequences, with feedback grounded in what they did. Simulations can also be deeply engaging, helping students connect with the material in new ways.
But simulations are also rare because they are very complex to design and build. The Generative AI Labs team has created and run educational simulations for more than a decade and has seen how effective they can be, as well as how difficult they can be to make at scale. Generative AI offers an opportunity to change this, democratizing the ability to make simulations while helping actually run complex simulations. It can sustain characters, improvise within a scenario, and adapt to each student’s choices in real time, which makes it possible to shift from scripting every path a student might take to designing the conditions under which learning should happen.
Primer is our attempt to build that system. It is open source, it is meant for educators rather than developers, and the educator’s judgment drives it: a structured interview elicits what you know about your lesson and your students, a design document makes it reviewable, and nothing is built until you approve it.
While we’ve provided information and explainer videos below on how it works, we’ve also produced a white paper on the Primer Project, that lays out in detail the concepts and architecture that drive it, including its prompts. All of this is released open source or creative commons so that other educators and researchers can build and expand on this foundation.

A Closer Look
An overview of the authoring and runtime workflow, from the educator interview to the student’s mentor, practice, and feedback sequence.
1. Primer Overview (2:21)
The whole arc, from an educator describing a lesson to a student practicing and receiving feedback. Start here.
How We Approached Using AI in Educational Simulations
Generative models are capable but not pedagogically reliable on their own. Left to their default settings, they help students too quickly, drift from the learning objective, and produce convincing roleplay without the structure that makes it educational.
Primer addresses this through structured orchestration: an architecture that organizes simulation design around principles of simulation pedagogy while accounting for the strengths and failure modes of large language models. The educator is the source of instructional judgment. A structured interview elicits that judgment, a design document makes it reviewable, and nothing is built until the educator approves it. Specialized agents then build and run the simulation, each working from inspectable artifacts rather than a single model’s memory.
Design Principles
Authoring by Conversation
An educator builds a simulation by engaging in dialog with an AI, not by writing code. The system asks the questions; the educator supplies the answers.
Educator as Expert
Instructors are the source of pedagogical and content knowledge. Primer elicits that knowledge, renders it as a reviewable design, and builds nothing until the educator approves it.
Learning Science in the Instructions
Scaffolding, retrieval, and feedback are not left to a model's general knowledge. They are written into each agent's instructions so they govern what happens during play.
Open by Design
Prompts, templates, orchestration patterns, and code are released together, so that educators, researchers, and developers can inspect every choice, test alternatives, and build on the result.
How Primer Works
Primer operates in two phases: authoring, in which the educator and a pipeline of agents design and build the simulation, and runtime, in which the student plays through it.
Authoring begins with a structured interview. The Intake Agent asks what lesson the educator wishes to simulate, then elicits the student/s educational level, prior knowledge, common misconceptions, along with the instructor’s teaching approach and assessment criteria, one question at a time.
Educators may speak or type, and may upload syllabi, lesson plans, or rubrics. Defining objectives and evidence of competency before any scenario is produced follows a backward-design framework. The interview yields a design document: learning objectives, scenario flow, characters, stakes, and adaptation logic, which the educator reads as a lesson plan, revises, and approves. Only then do the downstream agents act. These include:
- Intake Agent. Interviews the educator and records goals, student context, and source materials.
- Design Agent. Translates the interview into a structured design document for educator review.
- Casting Director. Converts the approved design into role-specific prompts for the mentor, practice, and feedback characters.
- Image Generation Agent. Produces the scene visuals that accompany the simulation.
- Experience Builder. Assembles prompts, images, and configuration into a playable experience the educator can assign to one student or a class.
At runtime, three student-facing agents guide the learner through preparation, practice, and feedback. A mentor activates prior knowledge and surfaces misconceptions. A practice agent plays the student’s counterpart in one of two modes. In Roleplay, the learner pursues a goal against a counterpart with motivations and constraints of its own. In Teachback, the learner explains a concept to someone who does not yet understand it and revises the explanation in response to questions. A feedback agent then acts as a coach, drawing on the transcript to discuss specific moments and how the learning transfers beyond the scenario. Each agent works on two levels: an in-world character the student encounters, and an instructional layer that tracks goals, calibrates challenges, and gathers evidence. After a session, the educator receives individual and aggregate reports on student performance.
Build on Our Work
Primer’s contribution is architectural: a working demonstration that educator intent can be carried from lesson design through student practice and feedback, and a modifiable foundation for research on educator-in-the-loop authoring, multi-agent educational workflows, and simulation-based pedagogy. Its prompts, templates, checklists, journals, orchestration patterns, and code are available on GitHub. Educators can refine characters and feedback criteria; developers can change models, orchestration, or storage, or add pedagogical templates beyond Roleplay and Teachback; researchers can study how educator input becomes runtime behavior. Because Primer is LLM-agnostic and runs with local or open-source models, it can be deployed under varied privacy and governance requirements.
The field has not yet settled how best to elicit educator expertise, apply principles of learning, or generate grounded feedback with these systems. An open system lets educators, researchers, and developers examine those choices together and build shared knowledge about what works.
Build Your Own
We have created a prompt that has an AI agent build its own version of Primer, keeping its pedagogy and teacher-centered design while letting the technical implementation vary. Give an agent the full paper and the prompt, and expect a long run: possibly several hours of agent work and paid model usage.
Primer is an experimental prototype, not a vetted classroom product. It has not been evaluated as a learning intervention under classroom conditions or reviewed for compliance with FERPA, COPPA, state student-privacy law, or district policy; institutions that use it with students are responsible for that review.
Primer was made possible through the generous support of the Bezos Family Foundation, the Walton Family Foundation, and the Charles and Lynn Schusterman Family Philanthropies.

