I Refuse To Build A Trap
Build with Attitude #5: How to product-think when AI builds at lightning speed
The typical vibe-coded tool starts with a business idea.
This one started with rage.
A professor on Reddit had built an elaborate trap for students: fake citations, trick questions, prompt injection.
The post radiated smugness. The comments cheered. And somewhere in the thread, the quiet part went unsaid: this person spent more time on surveillance than on teaching.
Dr Sam Illingworth read that post and felt an urge to fix this.
We’re educators, not cops.
Since then, that single line runs through everything Sam builds.
The Builder
On Substack, Sam is the writer behind Slow AI, one of the fastest growing publications in Education, where he’s been methodically unpacking what critical AI literacy looks like since mid-2025.
I connected with Sam when he first joined, for three reasons. One, he’s brilliant. Two, he’s humble. Three, we share a mission - helping people develop critical AI literacy.
What almost nobody knows: he’s also the editor of Consilience, a poetry magazine.
Poetry is where I think most clearly about voice, authenticity, and what makes writing human.
That combination matters.
Because the tool Sam built isn’t a tech product pretending to solve an education problem, but an education tool built by someone who’s spent years thinking about how language shapes power, and how power shapes blame.
Before Substack, before poetry, Sam was writing code. He authored over a million lines of IDL and Python. Then a decade of not coding at all.
When he came back to building in January 2026, the doubt hit fast:
I hadn’t coded for over 10 years since my first postdoc. Maybe I’ve forgotten everything. Maybe vibecoding is just for people who already know what they’re doing.
This is the part that “shipped in a weekend” posts rarely include: the moment before the first try, when you’re asking yourself am I even the kind of person who builds things?
You are.
The Product
The idea behind Sam’s tool, the Integrity Debt Audit, started with months of writing and thinking about a pattern he kept seeing in higher education: AI didn’t break assessments. AI exposed assessments that were already broken.
Assessment briefs that were already boring, generic, and disconnected from real learning are now exposed by AI. And educators panicking, policing, blaming students instead of fixing the actual problem: our curricula were never that good to begin with.
Students using AI isn’t the crisis. A brittle curriculum is. AI just turned the lights on.
Sam wrote about it again and again. Still, reading changes minds slowly. Using a tool can change habits.
The tipping point was that Reddit post. An educator designed a system of planted errors, fake citations, trick questions, and prompt injection to outsmart students. But outsmarting students is not the role of educators.
He’s spent more time on surveillance than on making the assignment worth doing. That’s when I knew this framework needed to be more than a Substack post. It needed to be a tool people could actually use.
The Integrity Debt Audit lets educators upload an assessment brief, and have it scored across 10 dimensions of vulnerability to AI. Instead of pointing out if students cheated, it points out where your brief is structurally weak, and suggests how to fix it.
This is a brilliant approach. Detection doesn’t fix anything, improving the briefs does.
And it’s brilliant for another reason too.
According to my research, The Integrity Debt Audit has zero direct competitors. Every tool in the space is fighting over AI detection (Turnitin, GPTZero, Winston AI, Copyleaks).
Nobody else has built a tool that audits assessment vulnerability to AI and tells educators how to fix it. Dr Sam Illingworth owns a category of one.
The Build
Sam’s tool choices:
Claude for coding
Github for source control
and Streamlit for deployment.
In the end, the real challenge wasn’t vibecoding itself. It was figuring out what anyone needed from it.
Early versions gave scores and that’s it. Users felt attacked. I had to add context, interpretation, framing:
‘This score is normal. Here’s why. Here’s what to fix first.’
The diagnostic data was never the point. Helping people feel capable of improvement was.
That last sentence. I want to underline it twice.
Once, because helping people feel capable is a mission Sam and I share.
Twice, because this is a product insight that goes way beyond education.
Every builder who ships a scoring system, a dashboard, a diagnostic of any kind should read this:
People don’t want data. They want to feel capable of improvement. Then they want to know where to start.
The Nos
What Sam deliberately left out tells you as much as what he built.
No comparison features
Would have been easy to build, would have driven engagement, but it turns diagnosis into competition. That’s the opposite of what I want. This isn’t about league tables. It’s about honest self-reflection without an audience.
No AI detection.
People kept asking: ‘Can it detect if the brief was written by AI?’ No. That’s the wrong question. I refuse to build surveillance tools.
No gamification.
AI kept suggesting badges and streaks. Sam’s response:
God, no. This isn’t Duolingo for assessment design.No corporate tone.
AI kept suggesting I write like a policy document. I kept refusing. The tool needs to sound human, warm, on your side. Not bureaucratic.
In one of my previous interviews, Jenny Ouyang refused gamification features. Karen Spinner rejected AI suggestions that didn’t serve her non-technical users.
Now Sam, same instinct.
Attitude isn’t what you ship. It’s what you refuse to ship.
The Real Constraints
Like Jenny Ouyang, Karen Spinner, and me, Sam builds in stolen time.
Kids, kids, kids! My young family are super fun, but they are FULL ON.
All my coding happens in stolen 45-minute windows after bedtime or before they wake up. No long, focused sessions. Everything had to be built in small, resumable chunks.
The part that may surprise you is that the constraint made the product better.
If I couldn’t understand my own code after a 3-day break, users definitely wouldn’t get it.
For non-product folks, that’s a design principle: forced simplicity as quality filter.
Where AI Helped, Where It Didn’t
Without AI, the Integrity Debt Audit wouldn’t exist.
I don’t have time to hand-code a full web app while managing research, teaching, and family. Vibecoding let me build something real in stolen hours. Without AI, this would be a Substack post and a half-finished Google Sheet.
But AI also tried to derail the build. Constantly.
Suggesting features I didn’t need. ‘Add user authentication! Add data visualisation! Add social sharing!’ Every suggestion sounded reasonable but would have tripled the scope. I had to keep saying no.
And the scoring prompts that “sounded good but were actually too vague.” Claude would generate wildly inconsistent scores until Sam tightened the rubrics.
What still doesn’t work:
The LLM scoring isn’t deterministic. Same brief can get slightly different scores.
Mobile experience is rough.
Non-English briefs aren’t supported yet.
The fact that Sam names all of this openly gains my immediate respect. No “it works perfectly” energy. No “shipped and done.” This is a builder who knows exactly where the seams are and isn’t pretending they don’t exist.
Sam’s Advice
Build the smallest version first and ship it scared. You’ll think it’s not ready. You’ll be right, it’s not. Ship it anyway. The feedback you get from one real user who actually tries your broken prototype is worth more than six months of planning. And honestly? Most people won’t even notice it’s rough. They’re just glad someone made a thing that helps them.
Three things worth underlining:
Ship it scared. I love this! I’m officially stealing this line from Sam, it’s brilliant.
One real user beats six months of planning. The feedback loop only starts when someone touches the thing.
People don’t notice the rough edges you obsess over. They notice whether it helps.
They’re just glad someone made a thing.
What We’re Learning
What if the problem isn’t students?
You can build the next AI detection wrapper. It’s easy. It’ll probably sell.
Or you can build with attitude:
Going one layer deeper and daring to ask hard questions
Recognizing that the problems usually have upstreams and downstreams
Saying no to useless feature suggestions
Building value for everyone involved, not just the buyer
I’m trying to help people understand what critical AI literacy actually is and how they can develop it without fear or blame.
How to Build with Attitude
What’s Being Built in the PwA Community Right NoW
Andreea Dalia Lazar, PhD built workshop canvas for product teams to help the assess the UX strenghts and limitations of new AI features.
Rebecca Spitzer built PocketHog, a native iOS app that puts your PostHog analytics on your iPhone home screen as live widgets.
Mia Kiraki 🎭 built a crossword to demonstrate a smarter way to build
AI workflows.
Raghav Mehra and Elena | AI Product Leader built BMAD,
the Breakthrough Method of Agent Development.
Justin launched a service for hand-drawn profile banners.
Soumya Sreeram built Carify, an AI‑powered car companion that brings clarity, confidence, and support into everyday driving.
Ibrahim Ayub built a lightweight motivational tool to do one good act
a day dailygoodbits.com.
Michael Sattler built a tool for aspiring novelists finish their first book:
Tanya built 7 mini apps to demonstrate that English language skills are your primary tool for building apps with AI.
theCR8Fconclave built ØRD1S, a healthcare intelligence platform
John Samuel is working on a new programming language that explores the intersection of vibecoding and natural language programming.
Jeremy Wright - Marketer/ECHO is working on an AI cognitive partner called ECHO.
Marcela Distefano built The Outcome Fairness Pack, a Notion template called designed to help evaluate AI in a more equitable and traceable way.
You Might Also Enjoy
Build AI Practice Loops So Students Get 5x The Reps Without 5x Your Grading by Doan Winkel
Are you tempted to start building but unsure where to start?
Start here:
👉 Vibecoding Resource Hub
👉 Vibecoding Guide From 2 Builders Who’ve Shipped
👉 Claude Hub: Code workflows, Cowork setups, Skills, CLAUDE.md files
👉 UI Elements Prompt Pack
Built something already? I’ll help it get seen.
You don’t need to do everything alone. Let our community amplify your work.
As a PwA member, you can list your projects on StackShelf App: an online directory I built so your projects are discovered by readers, fellow builders, and AI. As of today, StackShelf has redirected 5,201 unique visitors to creators’ product sites.
 | A guest post by
|
Woah, lovely collab!! 🥰❤️
“
Build the smallest version first and ship it scared. You’ll think it’s not ready. You’ll be right, it’s not. Ship it anyway.
“
absolutely love this advice. iterative deployments are so much better :)
nice collab!