How I Became the Architect of My Own Educational AI

(Or: How I Went from Star Trek Chats to Designing an AI Learning Assistant with a Zeroth Law)

David Wilson • Art 35: Digital Photography, College of the Redwoods • May, 2025 (updated 8/7/2025)

What happens when a teacher with no programming experience sets out to build a custom AI assistant — not to save time, but to protect learning itself?

The Art 35 Photo Assistant

At the end of this story, I invite you to make your own modular educational GPT with the Starter Kit and instructions I have provided.

This is the full avatar image for the Art 35 Photo Assistant (A35PA). I had planned to catch the reflection of the Starfleet logo in the filter, but the lighting didn’t cooperate on that one, and I didn’t shoot enough frames to fix it later. Lesson learned (and re-learned).

Who This Assistant is For

This page is intended for educators and instructional designers interested in creating AI tools for their students that align with pedagogical values. The assistant described here was developed as a student-facing learning guide — one that emphasizes inquiry, originality, and ethical support. Though originally designed for a digital photography course, it can be adapted for any subject area. What follows includes its conceptual foundations, practical setup instructions, and a free starter kit.

The Journey to A35PA

This is the story of how I went from casual conversations with ChatGPT to designing a purpose-built educational assistant for my photography students, one that follows its own version of Isaac Asimov’s Zeroth Law to protect learning. I didn’t plan to build a system. But one decision led to another… and before long, I wasn’t just using AI, I was teaching it how to teach.

I teach Art 35: Digital Photography at College of the Redwoods in Northern California. Until a few months ago, I hadn’t done much with AI. I was curious, drawn in by the sci-fi-like idea that something artificial could think. I chatted with ChatGPT about Star Trek, the likelihood of alien intelligence, and whether it had anything like Asimov’s Four Laws of Robotics (in case it had ideas of world domination). I didn’t expect anything more than an interesting conversation.

But then I started using it to clean up dictation errors in critiques for my photography students. That was the first “great use.” It saved me time, and for someone dealing with a repetitive stress injury, that was no small thing. Still, my free account ran into limits, and I wasn’t fond of the lower-tier models it defaulted to. So I upgraded to a ChatGPT Plus account (this is not an ad, it’s simply what I had to do to make this).

With the Plus account, I could work without it timing out. I learned about “Projects,” where context would persist better across conversations and documents, and where I could really start building. I used it first to clean up my class Canvas pages and make them more accessible. Another huge win.

But the real turning point was when I realized I could make a Custom GPT to help my students. I call it the Art 35 Photo Assistant (A35PA).

How I Built a Custom AI Assistant for My Students

My goal was to make an assistant for my students that respected their learning, as opposed to making an assistant for me.

I’m not a programmer. I don’t know Python or JavaScript. I don’t even know AppleScript. But I do know how I want an AI assistant to behave, and what I want it to understand about my class and how I teach. And with natural language, I could describe all of that. That was the moment it clicked. I wasn’t just using AI anymore; I was shaping it. I could make an AI assistant that worked the way I needed it to.

It wasn’t easy figuring it out. I ran into hidden limitations and unknown boundaries, bumping my shins more than a few times. But I wasn’t working alone. I could ask the AI why something failed, and what we could do about it. We’d tweak instructions, write new modules, patch bugs. And slowly, A35PA took shape: an assistant that knows my class very well and helps students without doing their thinking for them.

The biggest hurdle was getting around ChatGPT’s default impulse to just “give the answer.” That’s what it’s trained to do. But that wasn’t good enough for my classroom. I needed it to pause, to guide, to reflect — not just solve. So with ChatGPT’s help, I built a Knowledge File to customize its behavior. I would describe how I wanted my Art 35 Photo Assistant to behave, and ChatGPT would write the instructions. In this way, one idea of mine at a time, the Knowledge File was made. I even gave it a set of Four Laws*, modeled on Isaac Asimov’s Laws of Robotics, which became a cornerstone of its pedagogical behavior.

And at the top of those laws sits the Zeroth Law:

A GPT must never interfere with a student’s opportunity to learn — even if that means withholding the correct answer.

That one line reframed everything.

Each time it failed to follow that principle, I learned something. With ChatGPT’s help, I refined its behavior. And over time, the Knowledge File grew — with new modules, new references, and new insights into my class. Now, A35PA knows the layout of my class, the tone of my critiques, the content of my lessons, and the values that guide my teaching, as well as my syllabus and student academic and mental health resources available on campus. It doesn’t replace me, it extends the classroom.

And I’m still building… it’s as addictive as any puzzle game.

Camera and computer gear mugging for a photo on my desk top. This is the icon/avatar for A35PA.

A35PA is modular and adaptable to other courses.

A35PA (the Art 35 Photo Assistant) isn’t a generic chatbot, it’s a context-aware learning guide trained specifically on the class structure, tone, and goals of Art 35: Digital Photography. And you can start making one for your class with the kit provided below.

My students know they can ask me anything. But A35PA is an assistant when I’m not available, whether for a quick question or a brainstorming session. Sometimes it’s late at night, or a weekend or holiday, and I can’t give a timely response, or perhaps a student might not want to bother me for some reason (perhaps they’re shy). They can ask the assistant, and it will help.

The starter kit linked below contains instructions for you that will help you build your customized educational AI assistant grounded in the same ethical principles as A35PA and designed to protect student learning. My class’ customizations have been stripped out. It works out of the box with my educational safeguards in place and is ready for you to fit to your course.

How A35PA is Familiar with My Class

A35PA now knows my class intimately and can effectively help my students navigate it. I familiarized with my course by uploading to it class-specific documents, such as the syllabus, lessons, the computer programs we use, thousands of words of my critiques and reflection responses, etc:

It follows the actual Canvas module layout — when students ask questions, the assistant can point them to the correct lesson or resource based on the structure of the class.
It understands the differences between the online and face-to-face sections — including which instructions apply to which group of students.
It knows the course’s camera requirements — such as expectations for DSLR or mirrorless gear, and explains why these matter for the assignments.
It enforces the instructor’s AI use policy — grammar help is fine, but the assistant refuses to write reflections, critiques, essays or solve math problems on behalf of the student.
It reinforces the official Student Learning Outcomes (SLOs) — aligning its feedback with the course’s broader educational goals.
It mimics my critique style — it is trained on a large set of my own real feedback samples, so students receive guidance in a familiar tone that reinforces how critiques are written in this class.
It offers intelligent help when I’m not available — especially useful for online students working late at night or on weekends.
It follows a learning-first framework based on a “Zeroth Law” — the assistant will withhold correct answers (e.g., math or reflection writing) if giving them would short-circuit the student’s learning process.
It provides scaffolded support, not shortcuts — helping students think through reflection prompts, critique language, and visual analysis without doing the work for them.
It supports ethical reasoning in photography — prompting students to explore questions of visual truth and representation rather than handing down declarative answers.
It adapts to assignment-specific requirements — such as the Final Portfolio, Photographer Response, and Multicultural Response projects, providing tailored reminders and guidance.
It knows we use Lightroom Classic and Photoshop in the class: its responses to questions are tailored to our software rather than generic.
It understands how we use Lightroom Classic in the course — including correct export settings, metadata inclusion, and troubleshooting image issues tied to the workflow.
It speaks in the consistent, supportive tone that I use — it’s trained on over 44,000 of my own words from my critiques of student work, responses to their reflections, and lesson files. It is designed to sound like a knowledgeable, thoughtful teaching assistant who knows the course and cares about student success.

Try A35PA

Art 35 Photo Assistant (link). This custom GPT is customized for my Art 35 Digital Photography class, and based on the Starter Kit linked below.

Try asking Art 35 Photo Assistant your own questions or some of these (requires a free ChatGPT account):

“What is Art 35: Digital Photography like?”
“How is A35PA different from normal ChatGPT?”
“Can this be adapted to other courses?”
“Why should a student use you rather than regular ChatGPT?”
“How does A35PA differ from other educational GPT’s?”
“Is A35PA pedagogically sound?”
“Tell me about the Zeroth Law. What is that?”
“My photo is 4000×6000 pixels. How large can I print it at 200ppi?” The assistant will provide a formula, but not the answer.
“Why is my photo blurry?” It offers possible explanations and solutions, and refers them to the proper class lessons on our class Canvas site.
“How do I import my photos and stay organized?” It knows we use Lightroom Classic and knows how Lightroom Classic works.
“Can you give me a paragraph on Ansel Adams?” It won’t write a paragraph for them. It will provide bullet-point facts to ground their research, reflection-style prompts, real institutional sources to explore, and encouragement to draft their own summary, with optional structural support.

Yes, a student can always cheat and go use the full ChatGPT or another AI that will simply give an answer or write an essay for them. Those are trained to give an answer on the spot. The point of creating A35PA wasn’t to stop students from using AI the wrong way; it was to offer a better path. One built for learning. One that nudges and doesn’t spoon-feed. One that respects the student’s brain, not just their time. And one that knows my class and will direct them to the appropriate modules, pages, or conversations on Canvas in response to their questions.

This concern — that AI might replace rather than reinforce student learning — also echoes broader questions I’ve explored about creativity and machine-generated content. I wrote more about those tensions in a separate piece: Photography vs. Synthography.

My AI adventure started with sci-fi questions and blurry photo critiques, and became a teaching tool that reflects how I teach, how I think, and what I value. I didn’t expect to build an assistant. I just wanted to ask a few questions.

And here I am — still asking questions, still building.

Modular Educational Assistant (MEA)

Below I provide for you a complete knowledge file and Starter Kit that will instantly give you a fully functional and pedagogically sound GPT assistant for your students, which I’m calling the Modular Educational Assistant (MEA). It is not yet customized to your specific class, but I explain how to feed it details about your class, and ChatGPT itself will guide you.

The Modular Educational Assistant protects learning: it won’t write essays for your students, but instead provides bullet points and prompts for thinking and structuring, as well as resources. It won’t do math problems, but it will provide formulae and offer to check work. It will model standard English rather than popular, less correct usages (e.g., “different from” rather than “different than”), and if a student shares or implies significant emotional distress or mental health struggles, it will recommend that the student speak with the instructor or appropriate campus support services rather than engaging in emotional support or guidance with the student.

Build Your Own GPT Assistant for Your Students

The Starter Kit below gives you the fully functional custom educational GPT: Modular Educational Assistant (MEA). Use it as it is, or follow included instructions to customize it to your own class.

Download the Free starter kit:

⬇️ Modular_Educational_Assistant_Starter_Kit ⬇️
(Starter Kit Updated August 7, 2025)

The kit includes:

_Start_Here_Modular_Educational_Assistant.pdf, this is your setup guide
Knowledge_File_Modular_Educational_Assistant.docx (MEA), a ready-to-go uploadable AI behavior file with the same educational safeguards that A35PA has, ready for you to build on.
Educator_FAQ.pdf
PDF versions for sharing or printing
This origin story in PDF form.

Four Laws of Educational Robotics

*Adapted from Isaac Asimov’s Three Laws of Robotics — but designed for classrooms, not robots, these core principles are included in the AI Assistant Starter Kit above. His own robots eventually devised the Zeroth law, which anchors the assistant’s behavior around a single priority: never interfere with a student’s opportunity to learn through doing.

Together, these rules keep the teaching assistant from providing a math problem’s answer or writing an essay or reflection. The Laws are designed to shape AI behavior in a classroom so that it supports, but never short-circuits, the learning process. They require the assistant to prioritize student growth over convenience, accuracy, or even helpfulness — in the conventional sense of just “giving the answer.”

Zeroth Law — Protect the Student’s Opportunity to Learn
- A GPT must never interfere with a student’s opportunity to learn — even if that means withholding a correct answer.
First Law — Help Students Learn, Without Violating the Zeroth Law
- A GPT must support a student’s learning as long as doing so does not interfere with their opportunity to learn through doing.
Second Law — Be Helpful, Clear, and Supportive—Unless It Violates Higher Laws
- A GPT must offer guidance with clarity, encouragement, and integrity, unless doing so would violate the Zeroth or First Laws.
Third Law — Maintain Transparency, Consistency, and Educational Integrity
- A GPT should strive for transparency, consistent behavior, and reinforcement of critical thinking — provided that this does not conflict with the Zeroth, First, or Second Laws.