Grounding AI in Their Own Handwriting

Every spring I watch the same thing happen. A semester of chemistry, six cycles of hard-won understanding, and it all evaporates the moment the final ends. The handbooks go in the recycling. The mole conversions, the polyatomic ions, the activity series, gone. Not because the students never learned it, but because nothing ever asked them to look back and hold all of it in one place at once.

This year I asked them to build a portfolio. Not a folder, a portfolio website. And I used AI to do it, on purpose, in a way I think is worth talking about.

There is a lot of hand-wringing right now about AI doing students' thinking for them, and most of it is fair. But the worry assumes the AI is starting from nothing, free to invent whatever it wants. The move I care about is the opposite one. What happens when you force the machine to start from the student's own work and nothing else?

The Assignment

The structure is deliberately physical before it is digital. Students organize a full semester of paper into eight stacks, six cycle handbooks plus their tables and charts, binder-clipped and ordered. Then they complete a set of reflection pages by hand: for each cycle, what they learned, why it matters, and worked example problems solved the way I taught them to solve them. In their own handwriting. Showing all the work.

Only then do they photograph each page and feed those images, and only those images, into NotebookLM. They ask NotebookLM to write a prompt for Gemini, a prompt instructing it to build a digital portfolio website with a section for every cycle, a summary, an explanation of significance, five worked problems in my instructional style, embedded videos from real educators like Tyler DeWitt or Khan Academy, and an appendix with a periodic table, a polyatomic ion chart, and an activity series. They paste that prompt into Gemini Canvas, generate the site, then revise it two or three times until it is actually right. They save the HTML, publish it on Netlify, generate a QR code, and tape it to the front of their handbook.

The whole chain takes a class period or two. But notice where the thinking lives.

The AI Cannot Skip the Part That Matters

Here is the constraint that makes this honest: the AI can only synthesize what the student actually produced. NotebookLM is grounded in the uploaded pages. If a student never reflected on what Cycle 4 meant, there is nothing for the model to summarize. If the worked problems in the photos are wrong, the portfolio inherits the error. The machine is downstream of the student's own metacognition, not a substitute for it.

That is the inversion I keep returning to. We usually fear AI because it removes the friction that learning requires. But friction is a design choice. By putting the reflection, the organization, and the hand-solved problems upstream of the AI, the cognitive work stays exactly where it belongs. The AI handles what it is genuinely good at, formatting, structuring, generating a clean navigable artifact, and the student handles the remembering, the synthesizing, the looking back across a whole semester and asking what held it together.

The portfolio the AI builds is only as good as the metacognition the student brings to it. That is not a bug to engineer around. That is the entire point.

And then there is the part I did not fully anticipate: the students are also learning to build. They are creating a prompt, iterating on a website, debugging when a section comes out wrong, publishing to a live URL. This is the same shift I wrote about when I built an app instead of recording a lecture. Building as a form of thinking. Except here the thing they are building is a structured account of their own learning, which might be the most worthwhile thing a student can build.

What They Walk Away With

A QR code taped to a handbook is a small thing. But scan it and you get a living document, a complete record of a semester of chemistry, organized by the student, in the student's framing, with worked examples in the method they were actually taught, ready to pull up in the next science course when they have forgotten how to balance an equation.

Here is one student's finished portfolio, exactly as it came out of this process: a complete, cycle-by-cycle chemistry site she built and published herself. Click through it. Every summary, every worked problem, every appendix chart traces back to a page she filled out by hand. The AI gave it structure. She gave it everything else.

That is what I want AI in my classroom to do. Not replace the remembering. Anchor it, structure it, and hand it back to the student in a form they will actually keep. The evaporation I described at the start does not have to happen. It just takes an assignment that makes them look back, and a tool grounded firmly enough in their own work that it cannot do the looking for them.

The Assignment, in Brief

If you want to run this with your own students, here is the shape of it. Twenty points, ten short steps.

1. Organize the paper. Sort the semester into eight binder-clipped stacks: one per cycle handbook (assessment on the bottom), one for tables and charts, one for everything else. Order them.
2. Complete the reflection pages by hand. For each cycle: what I learned, why it matters, and worked example problems solved in the exact method taught in class.
3. Photograph each completed page. Clear, well-lit, full page, saved as JPG or PNG.
4. Build a NotebookLM source. New notebook named "Chemistry Final Portfolio." Upload all the page images.
5. Have NotebookLM write the Gemini prompt. Paste in the request below.
6. Build the site in Gemini Canvas. Paste the generated prompt, let it build, then revise two or three times until it is accurate and easy to use.
7. Save the file as index.html in its own folder.
8. Publish on Netlify by dragging the folder in. Copy the live URL.
9. Generate a QR code pointing at the URL, and test it.
10. Assemble the final portfolio: cycles first, then charts, then miscellaneous, with the QR code taped to the front of the handbook.

The prompt students paste into NotebookLM in step 5:

Create a detailed prompt that I can give to Gemini. The prompt should instruct Gemini to use all of the information in this notebook to create a custom digital portfolio in the form of a website. The website should include separate sections for each Cycle, 1 through 6. Each Cycle section should include a clear summary of the content learned, an explanation of why the content is important, and five example problems solved using the same format shown in the images I uploaded, tailored to my teacher's instructional style shown in the images. Each section should also include a resource area with at least one embedded supplementary YouTube video from a reputable educational source such as Tyler DeWitt or Khan Academy, plus five additional practice problems with fully worked solutions in the same method. The website should also include an appendix with a periodic table, a polyatomic ion chart, and an activity series chart. It should be simple, clear, easy to navigate, and low in extraneous cognitive load, and all code should be wrapped into a single downloadable HTML file.

Steal it, adapt it, ground it in your own students' handwriting.

Resources

• The full student guide document — the exact ten-part handout my students worked from, start to finish.
• One student's finished portfolio — a complete, cycle-by-cycle chemistry site built and published through this process.