Meet Marcus T.: Building an AI Agent That Individualizes Instruction for Missouri Teachers

A ninth grader reading at a third-grade level. A fifty-minute class period. No extra planning time. In a 60-minute workshop for Missouri K–12 teachers, that problem got solved live — with a working AI agent, standards-aligned content and the prompt behind it handed to every teacher in the room.

Ask a teacher what individualized instruction actually costs them and the answer is always the same thing: time they don't have.

The expectation isn't new. Meet students where they are, in the same room, on the same clock, with the same fifty minutes. Adaptive instruction software handles a version of this well — and carries a per-student license cost most districts can't absorb at scale.

So this workshop took a different route: rebuild the core loop — diagnose the gap, generate targeted instruction — out of tools teachers already have.

The room

Missouri K–12 teachers, 60 minutes. Shanice Brown co-presented with Josh Higgins, a former classroom educator — Josh carrying the classroom reality and the case study, Shanice the technical build and the live demo.

The audience self-reported a 2–3 out of 5 on AI. That number shaped the session more than anything else: fundamentals stayed brief, because the real gap wasn't understanding what AI is. It was knowing how to ask it for something useful.

Three levels of agent

Before evaluating an agent, teachers needed a working definition. The session used three levels:

Level 1 — it tells you. Level 2 — it makes it. Level 3 — it does it.

The build target was Level 1.

Naming the ceiling honestly mattered. Plenty of AI sessions imply Level 3 and deliver Level 1, and teachers walk away either disappointed or, worse, expecting the tool to do things it won't. Stating the limit up front made the thing that wasbuilt more credible, not less.

Meet Marcus T.

An agent stays abstract until there's a student in front of it.

The demo used Marcus T. — a composite persona based on a real former student of Josh's, with name and grade changed. A 14-year-old ninth grader reading at roughly a third-grade level. Six grades behind, entering high school. Lexile 480L. Sixty-eight words correct per minute. Comprehension gaps in inference and identifying main idea.

And the detail that made the room go quiet: he avoids written work, and the avoidance reads as anxiety rather than defiance.

Every teacher in that session has a Marcus. Naming him — with a reading rate, a Lexile score and a behavioral read — turned a software demo into a problem they recognized.

What the agent actually produced

The content behind the demo was grounded in real Missouri Learning Standards, not generic ones:

• 9-10.RI.1.A — inference and text evidence
• 9-10.RI.1.D — central idea and summarizing
• 9-12.LS2.C — high school biology, ecosystem dynamics

The instructional passage covered the Yellowstone wolf reintroduction and the trophic cascade that followed — real high school biology content, written down to a third-to-fourth grade reading level, paired with an eight-word academic vocabulary pre-teach.

That's the move worth noticing. Marcus doesn't get easier content. He gets the same content at a reading level he can enter through.

Then NotebookLM turned that passage into a podcast, flashcards, a mind map and a study guide. For a student whose strength is listening, an audio version isn't a nice extra. It's the access point.

The Hub, The Engine, The Output

The architecture was simple enough to redraw on a whiteboard:

The Hub — a Lovable site called "Classroom Ready in a Day." One URL. No deck, no screen-switching, no tab hunting mid-session.

The Engine — the master prompt.

The Output — the teacher-ready plan: lesson plan, worksheets and assessments, all exportable to PDF.

Teachers fill an intake form — grade, subject, standard, skill gap, prerequisite gap, class context, instructional time — and every generated plan follows the same specification: address the standard, a lesson plan per component, an example week of topics, a success metric for proficiency and example assessments in multiple choice, short written response and matching, with an answer key.

The form captures no student personally identifiable information. Student records fall under FERPA, and the design honors that directly. The form asks about the gap. Never about the student.

That constraint isn't a compliance checkbox — it's what makes the tool usable in a district at all.

The prompt is the product

The Engine is the actual intellectual property. The app is the wrapper.

That distinction did real work in this room. District AI policies vary enormously, and a prompt travels where an app sometimes can't. A teacher whose district blocks the Lovable site can still paste the prompt into whatever tool they're approved to use and get the same output.

So a "Take the Prompt" section sat directly beneath the agent, copyable, for teachers to replicate in their own projects immediately.

Teachers left with both the tool and the method behind it. Only one of those two things scales without a license.

Built for the room, not the demo reel

One design decision is worth flagging for anyone running a live technical demo: a "Load Marcus Example" button pre-loaded the entire scenario, so the walkthrough never depended on a working API call in front of a live audience.

Demo reliability over live improvisation. The teachers were there to see what the tool does, not to watch a loading spinner.

Why it landed

The throughline was the same idea underneath every AI literacy session in this portfolio: input quality drives output quality.

Teachers saw a vague prompt next to a strong one and felt the difference in the results. Then came the reveal — the agent performs that structured prompting for them, every time, automatically.

That's the shift the workshop was really teaching. From a vague ask and a disappointing answer, to a structured input and a usable plan.

What comes next

The Personalizer isn't a finished product. It's a pattern.

Swap the standards, swap the subject, change the intake fields, and the same Hub → Engine → Output architecture serves a different grade band, a different district or a different content area entirely.

The reusable asset is the prompt. And prompts scale without licenses.

Used well, AI makes it considerably harder for a student to quietly fall through the cracks.