Don’t Think Vibe Coding. Think Just-in-Time Modeling.

Source: Mike Caulfield Substack
Author: Mike Caulfield
Published: 2025-12-10
Source type: essay
Original source: https://mikecaulfield.substack.com/p/dont-think-vibe-coding-think-just

Private backup: the full article text is archived in the private repository at archives/articles/mikecaulfield-substack-com-dont-think-vibe-coding-think-just.source.md. It is not published on the public Quartz site.

Summary

Mike Caulfield argues that the educational significance of LLM-assisted coding is not software production alone, but just-in-time modeling as a way to think. Using Claude Code to investigate why Seattle’s earliest sunset occurs before the winter solstice, he builds and revises interactive models that test competing explanations about solar noon, latitude, time zones, orbital speed, and axial tilt. The article frames AI-generated models as learning objects that can expose gaps in understanding, invite verification, and make complex phenomena manipulable for learners who could not otherwise build the tools themselves.

Pull quotes

Model-first thinking

“Now, as I show below, anyone can spin up a custom model of the world (or in the article, the solar system) to better understand a problem or a phenomenon.”

Build to understand

“So I thought why not learn how this works by building a model? Why not have AI make a webpage that calculates and charts sunrise and sunset and shifting solar noon? Then I can look at those calculations and better understand what’s going on. And in the process of building the model I’ll better understand the phenomenon.”

Verification inside the build

“I should say that in narrating this process I’ve left out dozens of times where I’ve looked at numbers and asked if the numbers were right, or the model was right and once or twice I caught Claude in a mistake (for example, at first not shifting the rotation figure based on date) and most of the time learning how an assumption I had was not correct or didn’t take into account an unknown factor.”

Big ideas

• AI simulations need clear boundaries for learning
• Voice AI may make learning support easier to access

Claims

• AI-built models can help students show conceptual understanding
• Vibe coding is an early sign of broader knowledge-work change

Key evidence and examples

• Caulfield begins with a confusing AI-generated explanation of Seattle’s earliest sunset and realizes that apparent understanding collapses under closer questioning.
• He uses Claude Code to build a Flask-based interactive solar-time model without writing the code himself.
• Comparing cities such as Seattle, San Francisco, Houston, Bangor, Anchorage, and Santiago helps test explanations involving latitude, time zones, and orbital effects.
• The model helps separate day length from solar-noon shift, making the equation of time more visible.
• The learning value comes partly from checking the model and the AI’s assumptions rather than accepting either as authoritative.

Education relevance

This is directly relevant to AI literacy, STEM learning, computational thinking, and inquiry-based pedagogy because it treats AI as a scaffold for model-building and conceptual testing rather than as a shortcut around understanding.

Durability note

The specific Claude Code model-building example will date as tools change, but the educational pattern of using AI to build inspectable models for inquiry is likely to remain useful.

My notes