By Lew Ludwig
Markdown, project setup, and what to do when AI gives you garbage: this month, the tools.
"If I didn't know any better, I'd assume you joined last week's department meeting. Did you hide in the back corner when these conversations took place?"
That came in last month, after "I Worry." Blake, a POD colleague, was naming his own department's arguments back to me. Another message, from Steve, a friend who is a vice president at a construction firm: "I tried the project setup. It worked. Now what else can I do?"
Same column. Same readers. Two different asks. Two and a half years in, I'm noticing this is the pattern.
Blake isn't alone. Since “I Worry” ran, more of these messages have landed — colleagues writing back to say yes, that's the room I'm in, those are the conversations I'm having. They're not asking for an answer. They're asking for company. For confirmation that the struggle is real, that their concerns are shared, that someone else sees what they're seeing.
Steve isn't unusual. I also hear these things from my MAA crowd. Alissa wrote: "I love that you're willing to share so many of your ideas and activities for the rest of us to try using ourselves." Dana was more concrete: she's "eager to do the homework to get AI to assist with TILT a project/assignment this fall." Same posture, different specifics. They're past the wow, look what this can do phase. They want to know: how do I actually use this more effectively?
I've learned this column needs to do both. But this month, I'm digging into the tools. Not resonance. Not why should we engage? But how do we engage well?
Setup: Projects and Instructions (Quick Revisit)
Back in February, I wrote about making your hobbit-hole more comfortable. The idea was simple: setting up a project in Claude (or ChatGPT) where you store instructions, files, and prompts in one place. No more explaining context from scratch every conversation. Upload your syllabus, your assignment specs, your pedagogical frameworks. The AI has your "shelf" to reference. You still have to tell it where to look.
Some of you tried it. Some came back saying: "This works. Now what?"
That question — now what? — is what this month is about. Because a project is only as useful as the instructions inside it.
Most of us write instructions like an email: conversational, a little loose. Fine for a one-off. But if you're coming back to the same instructions week after week, you need them structured in a way AI can reliably understand and act on.
That's where markdown comes in. Not because it's fancy. But because it's a format both humans and AI read the same way. Once that clicks, you can ask AI to create markdown files for you, store them, iterate on them, build a library. You're not starting from zero anymore. You're building on what you've learned.
The Format: Markdown Files
If you've never encountered markdown, it's a way of writing plain text that contains clues about structure. A heading gets a #. Bold text gets asterisks. A list gets dashes or numbers. When you read it as plain text, it's still readable. But when a computer reads it, it understands the structure immediately.
Here's why that matters for AI: when you upload instructions in a Word document or a Google Doc, the AI sees all sorts of formatting noise. Fonts, spacing, colors. It has to filter through that to find the actual meaning. PDFs are even worse. They look clean to us, but to AI, they're a maze of layout instructions, embedded images, and formatting code. Reading a PDF burns through what AI calls tokens (its reading budget) and delivers messy results. It's the same reason PDFs are notoriously bad for screen readers and ADA compliance: the format prioritizes how something looks, not how it's understood.
With markdown, there's no noise. Just structure. The AI reads your instructions faster and more reliably. It knows what's a heading, what's a subpoint, what's an example.
You don't need to learn markdown syntax. Seriously. You just need to understand that it exists and why it's useful. Then you ask AI to create a markdown file of your instructions.
Project Prompt: Convert Instructions to Markdown
I've written some instructions for how I want you to help me with [my task]. Can you convert these into a well-structured markdown file? Use clear headings, bullet points for lists, and bold for key terms.
Feed it your instructions. Cut and paste, or upload a document. Within seconds, you have a markdown file. Copy the contents into your project's instructions so AI follows them every conversation. Save a copy on your computer so you can refine it over time and share it with a colleague. It's a living document, not a one-time thing.
Diagnosis: What Went Wrong?
You've asked AI to do something. It delivers garbage. Confidently wrong. Or just not what you needed.
Your first instinct might be to blame the AI. Sometimes you're right. But more often, you need to diagnose the problem. You need to figure out whether it's the prompt, the model, or your expectations that need adjusting.
Here's a simple framework: What went wrong?
The prompt was unclear. You asked for "an example of a difficult calculus problem" and got back something too easy, too hard, solving the wrong concept. The AI did what you asked — you just didn't ask clearly enough. Fix: rewrite the prompt with more specificity. "Create a related rates problem involving two variables changing over time, where students must set up implicit differentiation. The answer should require at least three steps."
The context was missing. You asked AI to "create homework problems on hypothesis testing." The AI delivered something, but the setup didn't match what your students actually need. Why? Because you didn't tell AI that your students confuse null and alternative hypotheses, or that they're working with real datasets, not textbook examples. Fix: tell AI about your students and their context. “I'm teaching introductory statistics. My students struggle to distinguish null from alternative hypotheses. Create five homework problems where they test hypotheses using data from real studies. Each problem should explicitly label the hypotheses before asking them to compute.”
The model has limits. You asked the free version of ChatGPT to write a proof by mathematical induction. It produced something that looked right at first glance but skipped the inductive step. Even if it had nailed it, you'd run out of free messages after a handful of follow-up prompts. The free tier isn't built for sustained, structured work. Fix: try a paid tier — the higher-end ChatGPT and Claude models give you both better reasoning and the room to keep working. Or recognize that some tasks are beyond what AI can do reliably and do them yourself.
Once you've diagnosed which bucket the problem falls into, you know how to respond. You're not starting over. You're making a surgical adjustment.
Maintenance: Tracking and Updating Instructions
You've built a markdown file. You've stored it in your project. You've used it a few times. What’s next?
The file doesn't stay static. Every time an instruction lands, notice why. Every time it misfires, jot down what you'd change. Then revise it. Don't wait for semester's end. The file evolves with you. You improve it as you go.
Keep a simple naming convention so you know what you're looking at. Something like calculus_homework_v1.md, then calculus_homework_v2.md when you refine. You don't need version control software. Just a clear system.
And here's the thing: these instructions get better with use. The first version is a draft. By version three or four, you've solved the problem of how do I ask for this. You've found the language that works. You've eliminated the noise. That's the version you share with a colleague. That's the version you build on.
Fellowship
Tolkien understood something about fellowship: not everyone travels the same way. Hobbits are practical. They notice details. They want to know how to get from here to there. Elves think in centuries. Men dream in empires. Each brings different gifts to the journey.
You're not one kind of reader. Some of you — Steve with his spreadsheets, Alissa and Dana with their assignments — want tools. Concrete, usable, buildable. Others of you, like Blake, want company. You're in those department meetings. You're hearing the arguments. You want to know someone else sees what you're seeing. Both are real work. Both matter.
This column does both. Some months, I'm giving you Tolkien and language for hard conversations. Other months, I'm giving you markdown syntax and error recovery. Two and a half years in, I've learned that's exactly what this community needs. Not one voice. Both.
So take what serves you this summer. Build something if that's what calls you. Sit with the questions if that's where you need to be. The work of engaging with AI, whether it's the work at your screen or the work in the meeting, matters either way. And that's the journey.
AI Disclosure: This piece was written in partnership with Claude, which helped me organize the structure, edit for clarity, and identify gaps. The ideas and experiences are mine.
Lew Ludwig is a professor of mathematics and the Director of the Center for Learning and Teaching at Denison University. An active member of the MAA, he served on the project team for the MAA Instructional Practices Guide and was the creator and senior editor of the MAA's former Teaching Tidbits blog. His new book, The Science of Learning Meets AI, co-authored with Todd Zakrajsek, was published in April, 2026.