Vibe Coding - round 2

Scaffolding is a key to transforming an AI project into a reliable, goal-directed product. Learn how our team revamped the "Burrow" project with engineering rigor, focused prompts, and a methodical approach, moving away from 'glamour apps' to proper product development at a much faster pace.

Vibe Coding - round 2
AI scaffolding is the foundation. Sidenote: chatGPT has issues generating images of laptops 😄

Scaffolding - n - the surrounding frameworks, architectural layers, and structured prompts that augment a language model's (LLM) core capabilities, transforming it into a more reliable, goal-directed agent capable of performing complex, multi-step tasks. 

I know the previous “eat your veggies” post wasn’t super fun, but a healthy AI development diet can help your product grow big and strong!

For round 2 of Burrow, the team (still just Christian and me) took a deep breath, and I started over with the presentation layer. Fortunately, Christian’s management of the pillars already had engineering rigor, so I had a good foundation to build on.

I created a new Burrow project in Replit and started with a focused prompt to design the interface with a more modest scope. The experienced product and business parts of my brain cautiously agreed to return, on the condition that they would monitor me closely and retreat into the shadows if I got too cowboy again. 

One thing I added to the initial scaffolding was Replit’s integration with Expo so that I could deploy natively to the web, iOS, and Android.

The next steps were:

  1. Setting up a church-and-state division between what Replit could modify and what Christian was working on. I had read-only access to the pillars he owned.
  2. Setting up Github. Replit uses a local repository automatically, but setting up a central or “remote” repository allows access for multiple developers - human and AI - to access the project, alongside all of the backup benefits.
  3. Setting up expertise-specific “agents” to give me more development rigor - more on that in a later article.
  4. VERY cautiously adding only the API endpoints I needed to get the MVP working, and stressing to Replit that it must prompt me before adding any new dependencies. In v1, I was adding 3rd party tools automatically without a care in the world, contributing to all that bloat.
  5. Focusing on getting the features on the screen working well and avoiding the urge to say, “it mostly works… let’s add more!”

The initial project creation was exhilarating - so much visible progress in such a matter of minutes. All of the following steps were methodical and less adrenaline-infused, but also felt great in their own way. The reward was building something that had value (hopefully) and quality, rather than a glamour app a la Homer Simpson.

This felt more like proper product development, just at a much faster pace.