What is programming-as-theory-building-skill?

This project is a plugin for Claude Code, an AI coding assistant, that tries to make the AI produce better code by grounding its work in a concept from a 1985 paper by computer scientist Peter Naur. The central idea from that paper is that when programmers build software, the most important thing they create is not just the code itself but their mental model of why the code is shaped the way it is: what rules it protects, what boundaries exist, and where future changes belong. The plugin tries to pass some of that discipline to an AI agent before it starts writing. The problem the plugin addresses is that coding AIs often write code that looks correct at first glance but misses the deeper reason a piece of logic exists. They may add a new helper function in the wrong place, write tests that pass without actually checking the business rule, or introduce abstractions that are not needed yet. This plugin adds a checklist of questions the AI must work through before making non-trivial changes: rebuild the mental model of what the code does, place new code next to the most similar existing code, keep changes small and focused, avoid adding flexibility that is not yet needed, and write verification tied to actual behavior rather than just syntax. The repository includes benchmark results comparing code generated by Claude Haiku with this plugin enabled, with a different plugin (by Andrej Karpathy), and with no plugin at all. Across the three prompt families tested, the theory-building plugin scored highest in weighted averages, with the biggest gains showing up in executability and test quality. The advantage was most visible with looser prompts where the AI had to infer more rules on its own. Installing the plugin is straightforward: it can be added through the Claude Code plugin marketplace with two commands, or installed manually by copying a single file into your Claude configuration folder. It also works with other AI coding tools by copying its rules into the appropriate config file for each tool. The benchmark data, raw results, and per-run notes are included in the repository under a benchmark folder for anyone who wants to inspect the methodology or reproduce the runs.