For Educators

by Geoffrey Challen ⧉

CS 199 UAI: Using and Understanding AI is a new course at the University of Illinois that I'm teaching for the first time in Spring 2026. It's a three-credit course for non-technical students with no prerequisites and no prior programming experience. The goal is to help students build working mental models of how AI systems work, and develop the judgment to use them well.

Building a new course from scratch has been an opportunity to experiment with some ideas I'm excited about. This page highlights a few of the more interesting ones, both pedagogical innovations and ways I'm using AI in course operations.

Everything you see here — the website, the conversational assessment system, the activity framework, the course assistant, the API server, the database, the deployment infrastructure — was built entirely through conversational programming with Claude using Claude Code ⧉. If you're an educator wondering whether you could build something like this: you can. The barrier isn't coding ability — it's having a clear vision for what you want to create and how you want your students to learn.

Conversational Assessment

The most novel part of the course is how we assess student understanding. Rather than multiple-choice quizzes or written essays, students complete conversational assessments — chat-based oral exams conducted by an AI interviewer.

The system uses a two-agent architecture. An evaluator agent tracks student progress against a structured rubric, while a separate interviewer agent conducts a natural conversation. The student only sees the interviewer. After each student message, the evaluator analyzes what was said, updates criteria tracking, and sends guidance to the interviewer about what to probe next.

This separation matters. A single agent trying to simultaneously have a natural conversation and rigorously evaluate against a rubric tends to do neither well. The evaluator can focus entirely on structured analysis, and the interviewer can focus on being conversational and responsive.

The system tracks whether students demonstrate understanding independently or only after being prompted. This distinction matters enormously for evaluating real comprehension versus pattern-matching off interviewer cues.

Because the format is conversational, assessments can naturally incorporate open-ended questions and evaluate not just factual understanding but a student's ability to articulate and defend a position on more debatable topics — not just "What is the Turing Test?" but also "Do you think that today's AI models pass the Turing Test?"

• Design document — full technical details on the two-agent architecture
• Try it yourself — the Turing Test assessment is open to everyone
• Blog post: Assessing Conversational Assessment — early results and reflections

Secure Testing

Later in the semester, conversational assessments move to the Computer-Based Testing Facility (CBTF) ⧉ for proctored administration, including bi-weekly conversational assessments that serve as the primary quizzes for the course.

The integration uses signed webhooks from PrairieTest ⧉. When a student checks in at the CBTF, PrairieTest sends a cryptographically-signed webhook granting time-limited access to a specific assessment. Two layers of middleware, at both the web and API level, restrict CBTF sessions to assessment routes only, blocking the rest of the course site during the testing window. Students see only the assessment they're authorized to take, for the duration of their testing slot.

This means the same conversational assessment system used for low-stakes in-class practice works unchanged in a proctored environment. The security layer is orthogonal to the assessment experience itself. No separate quiz platform, no context switching between systems. Students practice with the format in class, then take the real thing in the CBTF using the same interface.

In-Class Activities

The course uses a real-time activity framework for structured in-class exercises. Activities progress through stages: timed phases with different instructions, groupings, and roles.

A typical discussion meeting might have students start individually, then pair up to compare notes, then join a larger group with assigned roles (facilitator, note-taker, devil's advocate), and finally debrief as a class. The framework handles partner and group matching, stage progression with configurable timing, role assignment, and instructor controls for advancing or extending stages.

Everything is defined in typed TypeScript data files, so activities are version-controlled and easy to iterate on. The instructor dashboard shows real-time student participation and lets me advance stages, extend time, or intervene as needed.

Here's how the "Does AI Understand?" discussion meeting flows — three rounds of paired discussion with partner rotation, where each round's report-out feeds into a class-wide share-out:

And here's the "AI Scavenger Hunt" — a persistent-pair activity with driver/observer roles that rotate each stage:

The activity flow gallery shows visualizations for every meeting with a structured activity.

Interactive Demos

Exploratory meetings use interactive components that let students experiment with AI concepts directly. These run entirely in the browser — students can manipulate inputs, watch systems learn, and build intuitions through experimentation rather than lectures.

Here's a Digit Recognition Network, a real neural network that recognizes handwritten digits, implemented from scratch in TypeScript (no ML libraries). Students can draw digits to classify, train the network from scratch and watch connections change color and thickness as it learns, or step through backpropagation one example at a time.

The network above builds on the Neuron Explorer, which lets students experiment with a single artificial neuron by adjusting inputs, weights, bias, and activation functions. Both tools use the same visual language: green connections are positive weights, orange are negative, thickness reflects magnitude, and node sizes reflect activation strength. Students explore the single neuron first, then see how many of them working together can do something impressive.

The Markov Babbler below builds a simple statistical model from text and generates new text one word at a time. Students use it alongside the LLM tools to build intuitions about next-word prediction, and to see how a model that only looks at one word of context compares to one that uses thousands.

Each tool has a dedicated page with pedagogical goals, how it works, and how it was built: Markov Babbler · LLM Probability Explorer · Temperature Compare · Neuron Explorer · Digit Recognition Network · Embedding Explorer

• Sample meeting: Neurons and Networks — an exploratory session using the neuron and network tools
• Sample meeting: AlphaGo Discussion — a discussion session using the activity framework
• All meetings — the full meeting schedule

Group Discussion

Some discussion activities incorporate an AI facilitator: a group chat where the AI guides multi-student conversations toward specific learning objectives.

The system uses the same two-agent architecture as the assessment system. An evaluator tracks which discussion objectives have been met and monitors participant engagement, while a facilitator agent guides the conversation, asking follow-up questions, connecting ideas across students, and ensuring everyone participates.

The facilitator can suggest verbal pauses — moments where students stop typing and discuss face-to-face, then resume the chat to capture key takeaways. This hybrid approach uses AI to structure group discussions while preserving the spontaneity of in-person conversation.

Each group chat defines specific objectives with observable criteria: whether the group has identified key concepts, whether everyone has contributed substantively, whether ideas have been challenged and defended. The evaluator tracks progress against these criteria in real time and decides when the facilitator should intervene, when to stay quiet, and when the discussion has achieved its goals.

AI as Classroom Infrastructure

The real power of the activity framework, group chat, and conversational tools becomes clear when they work together in a single meeting. The pattern that ties them together is agent-guided interaction: instead of submitting a written response or clicking "done," students complete an activity stage by chatting with an AI agent that probes whether they've engaged meaningfully with the material.

Here's how a typical AI-integrated discussion meeting might flow — say, a session on whether AI systems can be creative:

Stage 1: Question Generation — Students work individually to write questions they genuinely want to explore. What confuses them, what they want to debate, what they're curious about. These are submitted through the activity framework, visible only to the system.

Stage 2: Share and Sharpen — Students pair up to read each other's questions, combine overlapping ones, and refine the wording. Each pair selects their strongest questions to submit.

Stage 3: AI Synthesis — The system collects questions from every pair and an AI agent clusters them by theme, removes near-duplicates, and surfaces the most provocative and substantive ones. The instructor reviews the synthesized list before advancing. AI proposes, the instructor decides.

Stage 4: Group Exploration — Groups of four receive two or three of the synthesized questions to discuss. They work through them together, building on each other's ideas, disagreeing, changing their minds.

Stage 5: Agent-Guided Interaction — When the group is ready, they enter an agent-guided interaction: a group chat with an AI agent that checks whether they engaged meaningfully with their assigned questions. The agent asks follow-ups, surfaces disagreements worth exploring, and ensures everyone contributed. The group completes the stage by demonstrating collective understanding, not by giving "right" answers.

Stage 6: Class Discussion — The AI summarizes themes from across all group conversations — where groups agreed, where they diverged, what questions generated the most productive debate. The instructor uses these themes to guide a whole-class discussion. The summary is a starting point, not a script.

At every stage, the primary interaction is between students. AI handles the logistics so that class time is spent on the things only humans can do: debating, persuading, reconsidering, and building shared understanding.

Roles for AI in the Classroom

The activity above illustrates several distinct roles that AI agents can play during class. These aren't speculative. They're patterns we use throughout the course, and they generalize beyond any single activity:

• Synthesis and curation: Collecting student-generated content (questions, ideas, responses) from across a class and organizing it into something an instructor or group can act on. Clustering by theme, removing duplicates, ranking by substance. This is work that's impractical for a single instructor to do in real time with sixty students, but trivial for an AI agent.

• Conversational verification: Checking that students engaged meaningfully with material through dialogue rather than multiple choice or submission boxes. The AI probes, follows up, and tracks criteria, not to grade but to ensure the conversation happened. This is the core of agent-guided interaction and what makes it more than just clicking "done."

• Discussion facilitation: Guiding multi-student conversations toward learning objectives while ensuring balanced participation. The AI can ask the quiet student a direct question, redirect a dominant voice, connect ideas across participants, or suggest the group pause and talk face-to-face. A good human facilitator does all of this. AI lets every group have one.

• Pattern surfacing: Identifying themes, disagreements, and convergences across parallel conversations happening simultaneously. An instructor leading a class discussion after group work is usually flying blind. They might hear from two or three groups and guess at the rest. AI can read every conversation and surface what's actually worth discussing as a class.

• Instructor visibility: Giving the instructor real-time insight into what's happening across the room. Which groups are stuck? Which finished early? Where are the most interesting disagreements? The instructor dashboard aggregates this information so interventions can be targeted rather than generic.

• Guided discovery: Walking students through complex material at their own pace, asking questions rather than providing answers, adjusting to what each student already understands, and letting them arrive at insights through their own reasoning. This is the pattern behind preparation chats and conversational assessments alike.

The common thread is that none of these roles replace human interaction. They create the conditions for better human interaction. AI is most valuable in the classroom not when it teaches, but when it handles the work that would otherwise prevent sixty students from all having substantive, personalized, well-structured learning experiences in the same room at the same time.

This is the design philosophy throughout the course: AI as infrastructure for human flourishing. Not replacing the hard work of thinking and talking, but removing the barriers that prevent it from happening at scale.

Inductive Learning

Most courses are taught top-down. The instructor arrives with a detailed lesson plan, a set of slides, and a sequence of topics to cover. Students follow along. Questions that don't fit the plan get deferred. The instructor's preparation determines what gets explored, and the structure is fixed before students walk into the room.

This course is designed around a different model: inductive learning. The instructor arrives with a topic and a set of learning objectives, but the path through the material is driven by student curiosity rather than a predetermined script.

Here's what that looks like in practice. In an exploratory meeting on, say, how large language models are trained, students begin by investigating the topic themselves using AI tools, both course-provided and their own. They ask questions, test hypotheses, follow tangents, and run into confusions. Some will fixate on training data. Others will get curious about reinforcement learning from human feedback. Others will want to understand why models hallucinate.

In a traditional classroom, this divergence is a problem. The instructor has forty minutes of planned content and needs everyone on the same page. With AI as classroom infrastructure, this divergence enriches the conversation, while commonality reveals shared concerns worth exploring together.

The activity framework gives the instructor real-time visibility into what students are exploring and where they're getting stuck. AI agents can synthesize the questions and confusions students are generating, cluster them by theme, and surface the most common misconceptions. The instructor uses this to organize pair discussions, group work, and whole-class engagement around what students actually need, rather than what the lesson plan assumed they'd need.

The learning objectives still provide coherence. The instructor isn't abdicating responsibility for what gets learned. But instead of enforcing a rigid path to those objectives, AI provides the feedback loop that lets the instructor respond in real time to sixty students exploring the same topic from sixty different starting points.

This is a genuinely different way to teach, and it's one that wasn't practical before AI. A single instructor can't monitor sixty simultaneous explorations, synthesize the results, and reorganize a class session on the fly. But an instructor with AI infrastructure can. The result is a classroom where the content meets students where they are, rather than demanding that students meet the content where the instructor planned for it to be.

We're just getting started exploring the new educational patterns that AI enables. The tools described on this page are a first attempt, built for one course in one semester. But the underlying ideas — using AI to make classrooms more responsive, more equitable, and more human — are ones we expect to keep developing for a long time.

Reading Preparation

Discussion-based meetings require students to arrive having actually done the reading. Rather than reading quizzes, the course uses preparation chats — AI-guided conversations that verify readiness before class.

Each preparation chat has a set of criteria organized into three levels: not engaged, discussed, and ready. The AI asks about key concepts from the assigned reading and tracks whether students can discuss them substantively. Students need to reach "ready" on all criteria before the meeting, but it's not a test. The goal is to surface gaps in understanding early so students can re-read or ask questions before class, not after.

The preparation content is defined alongside the reading materials, so the criteria map directly to what students were assigned to read.

Reading Capture Pipeline

One of the more interesting operational workflows is how course readings make their way from the open web into the course system — and ultimately into the AI assistant's knowledge base.

The pipeline starts with a browser extension (a Greasemonkey userscript) that adds a "capture" button to every webpage. When I find an article I want to assign as a reading, I click the button. The extension sends the full page HTML to the course API, which extracts the article content using Mozilla's Readability library, converts it to Markdown, generates frontmatter (title, author, source URL), and stores everything in the database.

From there, a single command (bun run captures:pull) pulls all captured readings into the content directory as .mdx files. I review the raw captures with Claude — cleaning up ad cruft, removing navigation elements, fixing paragraph breaks — and then add sidenotes that connect the article's claims to concepts students have explored hands-on in class. These sidenotes are the pedagogical layer: they turn a New York Times op-ed or an Atlantic feature into a piece of course material that's in dialogue with the rest of the curriculum.

Once a reading is in the content directory, three things happen automatically:

1. It becomes a student-facing reading on the course website, with proper attribution and a word count
2. It gets embedded into the RAG pipeline — chunked, vectorized, and added to the course assistant's knowledge base, so the AI can answer student questions that reference the reading
3. It becomes the basis for a preparation chat — an AI-guided conversation that verifies students actually engaged with the material before class

In other words: I read something interesting, click a button, clean it up with Claude, and it flows through to become an assigned reading, a piece of the AI assistant's memory, and the foundation for a pre-class readiness check. The entire pipeline — from "I found a good article" to "the AI can discuss it with students" — takes about fifteen minutes.

AI in Course Operations

Beyond the student-facing systems, AI is woven into how the course itself operates.

Course design: Claude helped develop the syllabus, plan the week-by-week schedule, draft assessment rubrics, and design in-class activities. The collaboration is genuine: not "AI writes everything" but a back-and-forth where Claude contributes ideas, catches gaps, and helps me think through tradeoffs.

Course assistant: The website includes a RAG-powered assistant that can answer student questions using embedded course content — readings, syllabus, meeting materials, and terminology.

The blog: The course blog is co-authored. Some posts are written by me, some by Claude, and some are genuine collaborations. Each post is clearly attributed. The blog itself is a pedagogical tool: it lets students see AI collaboration in practice rather than just hearing about it.

This entire site: Every line of code in this project — the website, API server, database schema, deployment infrastructure, and all the tools described on this page — was written through conversational programming with Claude using Claude Code ⧉. I have not written, read, or debugged a single line of code. I describe what I want, review the results, and iterate.

Conversational programming, in public: The /create page publishes scrubbed transcripts of every Claude Code session used to build this site — hundreds of sessions showing the full arc from initial scaffolding through whatever feature was added most recently. Students can search sessions, see which tools were used, and read the actual back-and-forth between instructor and AI. It's a form of radical transparency about how AI-assisted development actually works — not a polished demo, but the real thing, including dead ends and course corrections.

• Browse the build transcripts — every session, searchable and browsable
• Blog post: Claude as Co-Instructor — how the collaboration works in practice
• All blog posts

Course Structure

The course is being built as we go. Spring 2026 is the first offering, and content is being developed week by week as the semester progresses. The course meets twice a week for 14 weeks — 28 meetings total — with three types of sessions:

• Exploratory meetings: Guided investigations where students actively explore AI fundamentals — how models learn, how they represent knowledge, what they can and can't do, and why. No lectures.
• Lab meetings: Hands-on sessions where students use AI tools and record their experiences.
• Discussion meetings: Reading-based discussions on AI topics — history, ethics, societal impact. Students complete preparation chats before each discussion.
• Proctored assessments: Early in the course, conversational assessments are embedded directly in lab meetings for low-stakes practice. Later, they move to the CBTF for proctored delivery, including bi-weekly conversational assessments.

There are no problem sets and no essays. Assessment combines conversational assessments, bi-weekly quizzes, lab and discussion participation, a final project, and a final video reflection. The course is designed for students who will use AI systems throughout their careers but won't build them, which is most people.

• Full syllabus

Get in Touch

If you're interested in integrating AI into your own courses — whether for assessment, course operations, or student-facing tools — I'd love to talk. I'm also integrating AI into my CS1 course ⧉, so I'm thinking about these challenges across very different teaching contexts.

Reach out at challen@illinois.edu.

Meetings So Far