A small public school district in Washington state plans to walk away from about $200,000 to $250,000 in software subscriptions next school year. Not by laying off staff. Not by cutting programs. By having a handful of its own administrators describe what they need to an AI tool and ship the replacement apps themselves.
The district is Peninsula School District in Gig Harbor, Washington. The technique has a name: vibe coding. Andrej Karpathy coined the term in 2025, and on May 19, 2026, Karpathy started his new job at Anthropic, working under pre-training team lead Nick Joseph — the company that makes Claude Code, which is the exact tool Peninsula’s staff are using.
The loop between researcher, lab, product, and a small public-sector budget closed about as fast as a loop can close. That’s the actual story. Whether it survives contact with reality for two more school years is the real question.
What Peninsula actually built
The most-cited example is LessonLens. A teacher records a class, uploads the audio or video, and the app returns instructional feedback — for instance, that the teacher gave great task directions but didn’t pause long enough after asking a hard question. James Cantonwine, Peninsula’s director of research and assessment, built it because instructional coaches couldn’t give that kind of feedback at scale.
There’s a strategic-planning app the school board now uses. Kris Hagel, the district’s CIO, estimated it would have cost roughly $30,000 to $40,000 from a vendor. Cantonwine built it in a few hours.
The pattern matters more than any single app. Peninsula isn’t replacing a learning-management system or a student-information system. They’re picking off the “low-hanging fruit” — back-office workflow tools they decided weren’t as complex as the price tags suggested. The first confirmed casualty is Informed K12, a workflow automation product for HR and finance.
How the workflow actually works
Peninsula’s setup has three parts. None of them require a computer science degree.
| Component | Cost | What it does |
|---|---|---|
| Claude Code subscriptions | $200/month per user | Lets a handful of admins prompt the AI to build apps |
| AI Studio (private) | ~$600/month | District-hosted layer where the finished apps actually run |
| Provider agreements | Included | Data privacy deals with Anthropic, OpenAI, and Google for the underlying models |
A few district administrators have $200/month Claude Code subscriptions. Their finished apps run on a private gen-AI system Peninsula calls AI Studio, which costs roughly $600 a month and has data privacy agreements with Google Gemini, OpenAI’s ChatGPT, and Anthropic’s Claude. Within those agreements, personally identifiable information is stripped before any input reaches the providers.
That last detail is the one most write-ups skip past. Student data law in K-12 isn’t optional. A district that builds its own tools also takes on the legal responsibility a vendor used to absorb. Peninsula handled it by stripping PII before model inference. That works for some use cases. It doesn’t work for everything you might want to build.
What Karpathy actually meant by “vibe coding”
The phrase sounds breezy because it is. Vibe coding means describing what you want in plain English and letting AI write the code. You don’t read every line. You iterate by re-prompting. If the app’s broken, you tell the model what’s broken and it tries again.
It’s worth being honest about what this is and isn’t. A teacher describing a printable task-list generator and getting a working web page in fifteen minutes is real. A district CIO replacing an enterprise HR workflow product with one weekend of prompting is also real, but the testing burden quietly moves onto the district. Peninsula seems to know that. Hagel has been explicit: “You can’t just throw up any old thing that AI writes for you. There’s got to be some vetting going on behind the scenes.”
Why does the Karpathy-to-Anthropic move matter for any of this? Two reasons. First, the person who named the technique is now working on the next generation of the model that does the technique. Anthropic has said Karpathy will lead a team focused on using Claude to accelerate pre-training research — meaning the next Claude is going to be partly trained with help from the current one. Second, Karpathy has long said his real interest is education. “I remain deeply passionate about education and plan to resume my work on it in time,” he wrote in his announcement. Schools running on Claude Code today are a small data point on what that work might look like.
Where the $200,000 actually comes from
The savings figure makes for a great headline. The breakdown is more useful for anyone trying to copy it.
| Source of savings | Mechanism | Realistic for most districts? |
|---|---|---|
| Canceled SaaS contracts | Replace 3-4 niche tools with in-house apps | Yes, if the district has 1-2 people willing to prompt and test |
| Avoided new purchases | Decide not to buy a tool, build one instead | Yes, but requires changing procurement culture |
| Custom apps no vendor sells | Build things that didn’t exist as products | Yes, this is the biggest unlock |
| Reduced consulting fees | Skip vendor implementation services | Sometimes — depends on the tool replaced |
Peninsula has already identified three or four subscription tools it likely won’t renew. Most of those weren’t classroom-facing learning products. They were the boring back-office stuff that gets purchased because nobody on staff has time to build it. That’s the part of the ed-tech market that vibe coding actually hollows out first.
The catch nobody puts in the press release
This is where the article needs to slow down. The savings number is real, but it isn’t free.
“[AI-written code] appears to introduce more security vulnerabilities and bugs than a human would.”
— Torrey Trust, professor of learning technology at the University of Massachusetts Amherst
She’s right, and it gets worse in K-12: districts handle sensitive student data like IEPs and health information, which means any vibe-coded tool that touches personally identifiable information needs extra care.
There’s also a quieter point Trust makes that’s worth taking seriously:
“Most ed-tech tools are not designed by educators.”
Vibe coding doesn’t fix bad code; it shifts who’s writing it. That’s a win when the new writers are the people who actually live with the workflow. It’s a loss if the same staff now spend their evenings debugging things a vendor used to maintain.
Cantonwine himself isn’t fully sold. “I’m a little skeptical about using AI to write all of our curriculum,” he said. “I don’t see that happening in the near future.” That’s the right caveat from the guy doing the work. The apps Peninsula has built so far are tools for adults — coaching feedback, planning, workflow — not learning materials for kids. The distinction is the whole game.
The vibe coding tools districts are actually testing
Claude Code isn’t the only option. Other school IT leaders are looking at the same general category of products.
| Tool | Type | Cost range | Best fit |
|---|---|---|---|
| Claude Code | Terminal-based, code-aware | $200/mo Max tier | Complex multi-file builds; what Peninsula uses |
| Cursor | AI-powered code editor (built on VS Code) | $20–$200/mo | Staff who already know a bit of code |
| Lovable | Browser, non-technical | Free to ~$25/mo starter | First-time builders making simple web apps |
| Replit | Browser IDE with AI agent | Free tier + paid | Quick prototypes, classroom demos |
Cursor reached $1.2 billion in annual recurring revenue with a $29.3 billion valuation. Lovable hit a $6.6 billion valuation and processes around 100,000 projects per day. The category is real money now. Collins Dictionary named “vibe coding” one of its words of the year, which is usually the moment a term jumps from technical Twitter to school board meetings.
Why this scares the wrong people
The reflexive worry about AI in schools is “will it replace teachers.” Peninsula’s actual work suggests that’s the wrong fear, or at least not the immediate one. The teachers in this story still teach. What’s getting replaced is the middle layer of ed-tech vendors that sell districts $40,000 strategic-planning tools and $25,000 workflow automation subscriptions.
That’s a meaningful shift. The ed-tech procurement cycle has been slow, expensive, and weirdly disconnected from what teachers actually want for two decades. If a district CIO with a $200/month subscription can replace a four-figure annual contract in a weekend, the procurement cycle isn’t getting reformed — it’s getting routed around.
The people who should be reading Peninsula’s case study closely aren’t classroom teachers. They’re product managers at K-12 SaaS companies whose tools are most replaceable: HR workflow, scheduling, internal dashboards, simple form-collection apps. Hagel says he’s getting questions about the work from other district tech leaders — he was the center of attention at the Consortium for School Networking conference this year, fielding questions from people who want to do the same thing.
What I’d watch over the next 12 months
The honest version of this story is that one district has a chief information officer with a real engineering background, a small group of curious administrators, and a culture that lets them ship things. That combination isn’t standard. The reason I’d still take this seriously is that the tooling keeps getting cheaper and the prompts keep getting better. The bar to copy Peninsula’s setup is lower this June than it was in January.
The questions worth tracking:
- Will any of Peninsula’s vibe-coded apps break in a way that costs more than the savings? Likely yes for at least one of them. The question is which one, and whether the district keeps building anyway.
- Will another district announce a bigger number? Almost certainly. Wichita Public Schools’ CIO has been on podcasts with Hagel about exactly this approach, per EdTech Digest’s coverage of Rob Dickson and Hagel discussing it.
- What happens when a vibe-coded tool touches student data and something leaks? This is the failure mode the entire category is one incident away from. It will eventually happen somewhere. How that district responds will set the template for everyone else.
- Does Karpathy’s renewed education focus turn into a Claude-for-schools product? He hasn’t said so, but the alignment is uncomfortably clean.
What this means if you work in a school district
If you’re a teacher, the practical takeaway isn’t “learn to code.” It’s: pick one workflow that drains your week, write it down in plain English, and try it in Claude Code or Lovable on a Saturday. The bar to a working prototype is now hours, not months. If it falls apart, you’ve lost a Saturday. If it works, you’ve replaced a tool somebody was going to charge your district for.
If you’re a district CIO, the question isn’t whether to start. It’s how to put the right guardrails on it before something goes wrong publicly. Peninsula’s model — private AI layer, PII stripping, named owners for every app, explicit vetting — isn’t the only valid one, but it’s the closest thing to a working blueprint that’s been published. The districts that move next will either copy it or learn the same lessons the expensive way.
The $200,000 number is the hook. The real shift is that the people closest to the problem now have a credible path to building the solution themselves. Whether that’s a one-district fluke or the start of how K-12 software actually gets made for the next decade is a question with an answer roughly 18 months from now.
