Frank Hecker
2.9K posts

Frank Hecker
@hecker
Interests: education as talent development, the fate of artists in a log-normal world. Former work: @FrankHecker. Elsewhere: @frankhecker.com on the azure site.



One way we're upgrading our physics course is using manim animations for derivations. I want to show you how we go about this. So here's the topic that introduces the second kinematic equation. It consists of 5 short manim animations one after the other, starting with:



Making educational content is painfully slow. A single lesson with explanations, worked examples, and questions takes me a full day. I've been building workflows with Claude Code to cut that to 2-4 hours. Here's what's actually working: Key insight: examples >> instructions. The AI needs to see good lessons, not just read about what makes them good. It also needs the same context a human curriculum designer would: what prerequisites exist, how concepts were previously taught, what question formats work, what cognitive load looks like for this topic. So I've built a pipeline that can: Interact (safely) with my DB - First step was building a CLI for the database and knowledge graph, plus an export/import layer that dumps lesson content to markdown files. This is what Claude Code actually works on, and I review before it gives me the CLI commands to push back to the DB. Scope lessons - Interactive skill that helps define exactly what's testable and references previous lesson scopes to maintain consistency and manage cognitive load. Generate content - Separate skills for explanations, worked examples, and questions. Each pulls from exemplar lessons and the knowledge graph so it knows what to assume (and what not to) as well as how supporting lessons were taught. Create questions - The hard part. Multiple template-aware skills that can run Python to generate and validate coding exercises as well as LaTeX. Less prescription on format, more showing of past examples with clear guidance on the workflow to validate code runs. Run autonomously - An agent that bundles everything and checks its own work. I'm still iterating. The agent creates decent first drafts but I'm finding sequential skill-by-skill with human review beats full autonomy for now. Goal is to stop being the bottleneck on my own content pipeline. 😅 And however autonomous the generation gets, human review is always essential. Products that skip this are basically making QA the user's problem, not theirs.


@zenahitz I’ll bite on this. Let’s say the goal is something like “broad based scientific literacy” — the ability to understand the broad strokes of how the physical world works, as well as fluency in mathematical techniques you’d need to understand the average paper somewhere.











oooh this looks fun the cool thing is that under the hood courses are not a primitive in MA - it's all just a sea of individual topics in one giant directed graph - which reflects the fact that (esp in math) course boundaries are an illusion and all the knowledge is connected so an MA course (assuming you have developed all the necessary topics) is simply a curated playlist that picks out topics from the graph

Huge thank you to @HowardHughesHQ, @DavidOReillyHH, and Jim Carman. Alpha School’s #1 growth constraint isn’t demand. It’s real estate. To scale the best education, we need partners who aren’t just leasing space - they’re committed to bringing world-class schools to their communities. Howard Hughes gets it. Thrilled to bring Alpha School to The Woodlands in Fall 2026.









