Plumb: I built my friend a tutor for her own codebase, because "the AI writes the code now"

A software engineer friend recently posed a question that challenged the notion of continual learning, asking why she should continue expanding her knowledge when an AI agent could write code for her. Rather than providing a lecture, the reporter decided to build a solution to address her concern. The result is Plumb, a tutor designed for her specific codebase.

Plumb does not generate code itself, but rather ensures the user remains the person who comprehends it. To utilize Plumb, the user runs the tutor within their repository, which subsequently opens a chat interface. Prior to any modification, Plumb analyzes the real files, explains the intended changes, and outlines relevant design patterns. Users can express their preferences by marking items as favorites.

During the change process, Plumb engages the user with a series of questions about design choices, offering two to three reasonable answers, always excluding misleading options. The model also provides an "I don't understand" response, which is always an available choice. Following the change, the user can request a review with the /review command, which compares the code modifications against the decisions made earlier. Plumb identifies whether the changes were followed, altered, or not addressed at all.

Plumb also features a /tour command, which provides a guided tour of the repository, one file at a time, starting from an entry point. Users have the flexibility to skip any topic, and skipped topics will be revisited at a more convenient time, ensuring the user is not prompted unnecessarily. Plumb guarantees that every claim is referenced by its corresponding file:line, and it cross-verifies each citation with the actual file before displaying it.

Incorrect files, lines, or ranges are clearly marked on the screen. The model distinguishes between documented (commit messages or doc) and inferred decisions, and confirms user agreements explicitly. The Plumb model never presents a guess as a fact.

The user's knowledge, skipped topics, and decisions are saved in a .tutor/ folder as plain JSON files within their project, allowing them to open, edit, or delete these files as needed. The demo session demonstrates a real, unedited interaction on the public FastAPI full-stack template, showcasing Plumb's capabilities. The session employs the hosted gemma4:31b-cloud model via Ollama, with an alternative local Gemma option available for those with Laptop resources.

Plumb-cli is a local-first command-line tool that guides developers through understanding their own codebase and verifying their comprehension of architectural decisions before and after any code is written.