James Cantonwine has long wished teachers working in his district’s smaller schools of 250 students or less got the kind of personalized instructional coaching available to educators at the district’s larger campuses.
So Cantonwine, the director of research and assessment for the 9,000-student Peninsula school district near Seattle, used artificial intelligence to create the next best thing: LessonLens, which critiques teachers’ lessons based on recordings they upload to the tool.
LessonLens was inspired in part by a software demo Cantonwine saw that gave teachers feedback on the number of open-ended questions they asked in classes (which can spur creative thinking) compared with close-answered ones (where a simple “yes,” “no,” “true,” or “false” will suffice.)
“And I thought, that’s really cool, but [it] also seemed like it left a lot of value on the table,” Cantonwine said.
He believed AI could offer teachers more detailed feedback to help them reflect on instructional practice—without the feeling of judgment that comes with a human critique, even from a supportive coach.
Teachers can upload videos of their lessons to the tool, then select a set of guidelines against which they want to be measured.
Options include popular blueprints for best practice, such as Charlotte Danielson’s Framework for Teaching, Doug Lemov’s Teach Like a Champion, or Universal Design for Learning, as well as the district’s own instructional guidelines.
A biology teacher, for example, uploaded a video of her lesson on DNA to LessonLens and asked for feedback on how closely she adhered to what the district considers best instructional practice.
LessonLens complimented her for encouraging peer collaboration and providing clear directions but suggested she consider assessing how well individual students understood the lesson, as opposed to asking questions of the whole class.
One teacher pursuing her national-board certification asked LessonLens to evaluate a lesson she plans to submit to earn the credential.
Use of the tool is “totally optional,” and not part of a teachers’ performance evaluation, Cantonwine said. District leaders can see how many teachers used it but not who. And they can’t access the feedback the tool provides, though some teachers have shared it voluntarily to help Cantonwine test and refine the platform.
The lack of visibility for district and school leaders “was deliberate,” said Cantonwine, a former middle school science teacher who thought back to his own experiences with coaches in designing the tool.
“One of the big challenges for any instructional coach is: How do you make sure that the teacher understands that you’re not big brother? That you’re not there to spy on behalf of the principal, but that you’re there to help a colleague get better?” Cantonwine said. “And so, if we’re going to have an AI tool help with coaching, it’s got to be able to do the same thing.”
Vibe coding can be problematic but can pay dividends
Cantonwine developed the tool using Claude Code, a widely available AI coding application. (Competitors include: Codez, Cursor, Replit, and Loveable.)
The approach, known as “vibe coding,” allowed the district to create a coaching platform that’s more customized—and cheaper—than anything available on the open market.
Tools created using vibe coding can be glitchier and more prone to security vulnerabilities than those coded by humans, experts say. But district leaders with significant computer science expertise may be able to alleviate those problems.
Cantonwine doesn’t have that sort of background, but he’s collaborated closely with Kris Hagel, the district’s chief information officer, and other staffers with deeper coding experience.
The district has created a range of other tools using vibe-coding. One is a tool that can comb the internet for scholarships that Peninsula students may be eligible for and add the information to district communications and postsecondary planning kits for kids and families. Another is an app to help parents compare Peninsula’s academic performance with that of other Washington state districts with similar demographics.
Cantonwine knows LessonLens has a few flaws. For instance, when one teacher submitted an audio file of her lesson, as opposed to a video, the tool told her that she needed to stand closer to students or move about the room more.
The teacher had, in fact, used that “proximity” strategy, but LessonLens wasn’t able to pick up on that when it had no visuals to analyze and made the critique without the relevant information.
What’s more, as far as Cantonwine knows, LessonLens has mostly been used by some of the district’s instructional high fliers.
He’s not sure yet how the tool would handle a lesson that truly flops—or whether a teacher would even want to record and share their failures, no matter how much privacy they’re guaranteed.
“I don’t think there’s an authentic way for me to grab a sample from a teacher who’s really struggling in the class in the way I kind of wish I could,” in order to test how the tool would respond, Cantonwine said.
But even if LessonLens needs refining, he’s excited about the tool’s ability to provide “a very Peninsula-specific version” of PD.
“It’s not something we’d be able to buy off the shelf,” Cantonwine said.
