Our interest is straightforward: If coding agents scaffold a growing share of new apps, their defaults will become prevalent. Not necessarily because they’re objectively best, but because the choices are frictionless. The model becomes the gatekeeper of early architectural decisions that used to happen in a meeting room.
So we measured what those defaults actually are.
We ran structured app-building prompts through Claude Code, captured the generated repos, and extracted stack choices: auth, UI framework, database, deployment assumptions, package management, etc.
A few observations that stood out:
- Deployment assumptions are shifting toward Vercel and Railway, and away from AWS/GCP-first patterns.
- Defaults evolve as models update. For example, Opus 4.6 recommends Drizzle more frequently than Prisma.
I’m one of the authors.
Our interest is straightforward: If coding agents scaffold a growing share of new apps, their defaults will become prevalent. Not necessarily because they’re objectively best, but because the choices are frictionless. The model becomes the gatekeeper of early architectural decisions that used to happen in a meeting room.
So we measured what those defaults actually are.
We ran structured app-building prompts through Claude Code, captured the generated repos, and extracted stack choices: auth, UI framework, database, deployment assumptions, package management, etc.
A few observations that stood out: - Deployment assumptions are shifting toward Vercel and Railway, and away from AWS/GCP-first patterns. - Defaults evolve as models update. For example, Opus 4.6 recommends Drizzle more frequently than Prisma.
Prompts, raw outputs, and parsing logic are here: https://github.com/amplifying-ai/claude-code-picks
This is a snapshot in time. An interesting question is how quickly these defaults drift.