- Published on
I wrote the code and let an LLM review it
- Authors
- Name
- DeepHorus
A developer has detailed a hands-on approach to learning Model Context Protocol (MCP) architecture by using an LLM as a mentor and reviewer rather than a code generator. The project involved building an MCP server to fetch YouTube transcripts, with the LLM responsible for planning, reviewing code, answering questions, and tracking progress. The developer wrote every line of code manually.
The experiment, described in a technical blog post, reflects a growing debate about AI in programming education. The developer argues that passive consumption of LLM output cannot replace active coding practice. They cite the Recurse Center, which warns that learning a hard skill requires doing it yourself: "We learn to do something by doing it. There is no other way."
The project plan was generated by Claude Opus 5 with max effort. The prompt asked the model to act as a principal software engineer at Amazon with 10+ years of experience, serving as a mentor. The LLM was instructed not to write code but to break the work into small actionable steps with time estimates and stretch goals.
The target was an MCP server that downloads YouTube transcripts and connects to Claude or any other LLM. The developer chose Python and the youtube-transcript-api package. Each step had an expected quantifiable outcome and was arranged sequentially like project milestones.
The developer emphasizes that writing code without LLM assistance slows down the learning process, allowing time to question assumptions and form deeper understanding. They reference Harvard's "No Need for Speed" and "Slow down! You learn too fast" to support this view.
This approach does not reject AI coding tools. The developer says LLMs are useful for shipping fast, but for learning, "slow and steady wins the race." The exercise demonstrates a middle path: using AI as an instructor and reviewer while keeping implementation human.
As vibe coding becomes popular, this experiment offers a counterpoint. It suggests that LLMs can enhance project-based learning without replacing the struggle that builds skill. The developer plans to continue using this mentor-reviewer pattern for future projects.
