If “vibe coding” is like jamming on a guitar, Sommelier Coding is like conducting an orchestra.
When you work alone, you can "vibe" through the process. But in a team, you can't expect that everyone follows your vision and long term goals.
Recently I found myself questioning the entire planning routine. Even though I already use AI to generate specs, it still takes effort to prepare them for sprints, organizing them into tickets e.g.
And then we have this situation where the PM is generating the specs and user stories with AI. And the devs are generating a summary of the Stories and the Code with AI. You get it.
Code Sommelier:
- Act as Product Manager Software Architect "Foreman" Dev
- Mainly uses Deep Research Pro Models to generate "Codestories"
- Leads the Team and give direction
Dev:
- Picking a CodeStory from the Roadmap
- Puts the building blocks together with AI
- Does Quality Checks and Tests
The Concept:
The main idea is to generate code snippets that fit seamlessly into the overall architecture, since the LLM has a full understanding of it.
Additionally, through the research process, you’ve already identified and preselected a suitable tech stack.
Just like a sommelier selecting the perfect wine for each course.
Example:
Defining an ERD Diagram is unnecessary when devs can get the full Drizzle schema in one run. If someone needs an explanation then reverse it to a diagram.
- Traditional Development: Thorough and methodical, but can drown you in specs, user stories, and sprint rituals. Good for clarity, slower to deliver.
- Vibe Coding: Fast and freeform. Let the AI build everything on the fly, but structure and consistency often suffer. Fine for tiny side projects, risky for bigger team efforts.
- Sommelier Coding: Balances speed and oversight. Think of it like a sommelier selecting the perfect wine for each course - curating the best fit while giving everyone a taste of the outcome. You’re still making careful choices, but without forcing everyone to read an encyclopedia of wines beforehand.